Professional Documents
Culture Documents
DRAFT
USERS : PLEASE PROVIDE FEEDBACK AND COMMENTS TO Joy de Beyer ( jdebeyer@worldbank.org) and Ayda Yurekli (ayurekli@worldbank.org) World Bank, MSN G7-702 1818 H Street NW Washington DC, 20433 USA Fax : (202) 522-3234
Contents
DRAFT....................................................................................................i
I. Introduction
Purpose of this Tool...............................................................................................1 Who Should Use this Tool....................................................................................2 How to Use this Tool.............................................................................................2
The Reason for Analysis of Demand.....................................................................4 The Economic Case for Demand Intervention......................................................4 Analysis of Demand for the Policy Maker............................................................5 Design an Analysis of Demand Study...................................................................6 Components of a Study...........................................................................6 The Nature of Econometric Analysis......................................................7 Resources Required.................................................................................7 Summary................................................................................................................8 References and Additional Information................................................................8
9 11
Choose the Variables...........................................................................................11 Data Availability...................................................................................11 Data Types............................................................................................12 Prepare the Data..................................................................................................13 Data Cleaning and Preliminary Examination........................................14 Preparing the Data Variables................................................................14 References and Additional Information..............................................................19
20
Determine the Identification Problem.................................................................20 Test for Price Endogeneity....................................................................21 Find Instrumental Variables..................................................................21 Select the Demand Model Type..........................................................................21 Select the Functional Form..................................................................................22
24
Demand Specifications for Annual Time-Series Data........................................24 Dependent Variable...............................................................................24 Conventional Demand Model...............................................................26 Myopic Addiction Demand Model.......................................................26 Rational Addiction Demand Model......................................................26 Double-Log Functional Forms..............................................................27 Semi-Log Functional Forms.................................................................27 Demand Specifications for Quarterly Time-Series Data.....................................28 Conventional Demand Model...............................................................28
ii
Economic Analysis of Tobacco Demand Myopic Addiction Demand Model.......................................................30 Rational Addiction Demand Model......................................................30 Demand Specifications for Monthly Time-Series Data......................................30 Apply Instrumental Variable Techniques............................................................31 Administer Specification and Diagnostic Tests..................................................33 Coefficient Tests...................................................................................33 Residual Tests.......................................................................................34 Specification and Stability Tests...........................................................37 Multicollinearity....................................................................................38 Specification and Diagnostic Test Examples........................................38 Cross-Sectional Data...........................................................................................39 Understand the Limitations of Aggregate Time-Series Data................40 Investigate Individual-Level Demand Decisions..................................40 Define Key Variables............................................................................41 Specify Demand Functions for Cross-Sectional Data...........................46 Demand Models for Household or Individual Level Data....................47 Expect Results for Quantitative Independent Variables.......................51 Expect Results for Qualitative Independent Variables.........................52 Pooled Time-Series and Cross-Sectional Data....................................................55
56
Calculate Elasticities of Demand.........................................................................56 Conventional Demand Model...............................................................56 Addictive Demand Models...................................................................58 Expect Results for Quantitative Independent Variables......................................59 Price......................................................................................................59 Income...................................................................................................61 Advertising and Promotional Activity..................................................61 Nicotine Addiction and the Role of Past and Future Demand..............65 Time Trend............................................................................................67 Expect Results for Qualitative Independent Variables........................................68 Health Information and Counter-Advertising....................................68 Smoking Restrictions............................................................................69 Include Other Independent Variables..................................................................70 Prices of Complements and Cross-Price Elasticity of Demand............70 Prices of Substitutes and Cross-Price Elasticity of Demand.................71 References and Additional Information..............................................................71
VIII. Another Demand Model: Error Correction Models and Diagnostic Tests
73
Definitions...........................................................................................................73 Cointegration.........................................................................................73 Collinearity............................................................................................73 Error-Correction Model........................................................................74 Johansen Cointegration Procedure........................................................74 Price Endogeneity.................................................................................75 Stationarity............................................................................................76 Assumptions and Requirements..........................................................................77 Prepare for Regression Analysis.........................................................................77 Apply Tests for Non-Stationarity and Cointegration, and Specify ErrorCorrection Models...............................................................................................79 Non-Stationarity and the Problem of Spurious Regression..................79 Test for Non-Stationarity......................................................................80 Cointegration Relationships..................................................................82 Cointegration Testing: Estimate the Long-Run Demand Relationship.83 Estimate an Error-Correction Model and Short-Run Demand Relationship..........................................................................................86 Non-Stationary but Cointegrated Variables..........................................87
iii
88
Understand the Objectives of the Dissemination Phase......................................88 Identify the Composition and Requirements of the Target Audience.................89 Audience Sub-Groups...........................................................................89 Information Requirements of Audience Sub-Groups...........................90 Handling Research Results and Policy Implications............................92 Structure the Research Reports...........................................................................93 The Internal Analysis of Demand Report..........................................93 The External Report on the Economic Study of Tobacco Control Issues.....................................................................................................94 Explain Research Results....................................................................................95
X. Additional References
96
iv
Figures
Figure 3.1 The Advertising Response Function..................................................64
I. Introduction
The tobacco epidemic is a worldwide phenomenon with significantly destructive effects on developing, transitional, and industrialized nations. The first scientific evidence on the health consequences of tobacco consumptionspecifically, smoking was discovered in industrialized nations. As a result, the economic analysis of tobacco control issues began and was developed in these countries. Due to the origin of such research techniques, the English language literature is dominated by contributions from the United States. However, a new generation of economists and other analysts in low- and middle-income countries is developing programs of research into economic issues around tobacco control tailored to their own particular situations. This tool, along with the others in this series, is intended to assist such research initiatives.
products causes disproportionately far more disease and death than chewing or inhaling tobacco. By contrast, in developing countries cigarette substitutes (in the form of other smoked and smokeless tobacco products and non-tobacco smoked substances) have much greater importance in local markets. Given such mixed market importance and use of cigarettes and non-cigarette products, this tool attempts to focus on demand for tobacco products as a whole, making allowance for differences in category of tobacco products (or their substitutes or complements) where necessary. For both health and economics reasons, however, tobacco control initiatives should continue to focus primarily on the consumption of smoked rather than smokeless tobacco products for the foreseeable future.
the identification problem, demand model types, and functional form. The second step is addressed in the Specify the Demand Function chapter, in which options include conventional and addictive models for aggregate annual, quarterly, and monthly time-series data. This chapter also summarizes salient issues involved in econometric analysis of demand using cross-sectional and pooled time-series of cross-sectional data. Meaningful interpretation of the demand analysis results is urged in the Review and Understand the Results chapter. This chapter outlines the calculation of elasticities of demand, suggests the expected nature of estimation results for each of the key variables, and explains these expectations in terms of prior research into the nature of the variables and the mechanisms involved in their impact on demand. The chapter Another Demand Model: Error Correction Models and Diagnostic Tests briefly covers variable stationarity, cointegration, and the use of error correction models, alternative regression techniques, and specification and diagnostic tests. This chapter also contains definitions and background information with which every reader should become familiar. Further, this chapter includes the assumptions made by the author, as well as the requirements expected of the reader to best utilize this tool. The conclusion is presented in the Disseminate the Research Findings chapter, which covers the presentation of results and their policy implications in research reports, structured specifically to suit the information requirements of different audience sub-groups. Readers interested in additional research and empirical studies on demand analysis should refer to the Additional References chapter.
Before commencing an analytical study of the demand for tobacco products, it is essential to define clearly the purpose and objectives of the research, plan the analytical process, and ensure that adequate resources are secured for the study. Only a brief overview of these tasks is addressed here. Refer to Tool 1. Political Economy Issues for detailed coverage of analytical studies and their importance in tobacco control efforts.
1. There is information failure about the health risks of smoking. Because the tobacco industry has concealed and distorted information on the health risks of smoking, and because there is a delay between starting to smoke and the onset of tobacco-related disease, consumers tend to underestimate the health risks involved. While this underestimation is particularly prevalent in low- and middle-income countries, consumers in all countries may not grasp the scale of the health risks of smoking, even when they have been informed of them, and may not apply this knowledge to themselves. 2. There is information failure about the addictive nature of tobacco consumption. Smokers acquire psychological addiction (habit formation) to the act of smoking, and physical addiction to nicotine. Physical addiction in particular means that the effort and discomfort involved in quitting smoking are significant. Many prospective smokers, and particularly adolescents, underestimate the risks of becoming addicted to nicotine, and once addicted face high costs in trying to quit. These two information failures result in high private costs of death and disability for smokers. 3. Smoking imposes external costs on non-smokers. Direct physical costs to non-smokers include the health impacts and nuisance value of environmental tobacco smoke (e.g., passive smoking) and the greater risk of fire and property damage. Financial costs borne by people, whether or not they are exposed to tobacco smoke, include tobacco-related public health care costs and cross-subsidization of tobacco-related private health care costs. In addition, caring externalities include the emotional suffering of non-smokers due to the illness and death of smokers. In sum, the existence of ignored internal costs (in the form of harm to smokers themselves) and external costs (in the form of harm to others) justifies both government intervention and research on the effects and benefits of alternative policies to limit demand of addictive substances such as tobacco. Analysis of demand for tobacco products is a crucial component of such a research program.
of demand for tobacco products is not only that these factors (among others) influence an individuals propensity to consume a particular tobacco product, but that they influence the propensity to consume at a particular price. In other words, demand factors influence the price-responsiveness of consumers, and a major role of analysis of demand for tobacco products is to investigate and explain how and to what extent this price-responsiveness is influenced by which demand factors. Thus the role of an analysis of demand for tobacco products is to qualitatively explain the relationship between certain demand factors and the price-responsiveness of consumers and their demand for tobacco products, and then to quantify that relationship using econometric techniques. The estimates of the parameters of demand for tobacco products can be used to predict the direction and degree of impact on demand of such control measures that: increase the excise or ad valorem tax on tobacco products;
impose restrictions on smoking in public places and private workplaces; regulate tobacco product advertising; restrict sales of tobacco products to minors;
develop and disseminate information on the health risks of smoking through counter-advertising campaigns; and provide access to smoking cessation programs and nicotine replacement therapy.
The information from these predictions can be provided to policy makers and others concerned with tobacco control.
A data preparation phase to: gather background information and the data to be used for detailed analysis; evaluate and clean the data; and transform the data. An econometric analysis phase in which to:
3.
if necessary, correct the model specifications and perform the analysis again. 4. A dissemination phase in order to communicate the results and findings of the study to all interested parties, particularly those organizations that originally commissioned the study.
There is a need for pragmatism in designing the demand analysis. For example, compromises may be made regarding the inclusion of variables in the demand specification, or the quality and availability of the data required may be balanced against the likelihood of specification error resulting from omitting a significant variable. In such instances, a researchers experience and clear-headed judgement is particularly important.
The analysis process should be documented in detail. This helps researchers cope with changes in data or estimation technique that are forced upon them by measurement or specification problems. It also lets them confidently answer questions from analysts, policy makers, and lobbyists regarding the data used, the econometric techniques applied, and problems they encountered and solved.
Resources Required
The resources required for the study will include the following: an economics graduate with competence in, and some applied experience of, econometric analysis;
an up-to-date econometrics software package (such as SAS, STATA, or SPSS) that can, for example,
handle relatively recent innovations such as the Johansen cointegration procedure; data on tobacco product demand variables, in sufficiently numerous observations and of sufficient accuracy, to meet the analytical requirements of the study; and
an adequate budget of both time and financial support to guarantee the effective use of the above resources.
Summary
Below is a list of questions and concerns to address that can act as a checklist for researchers, analysts, and policy makers initiating economic analysis of demand for tobacco products. With the aid of this tool, the answers to these questions should be clarified and formalized during the planning of the study, and then communicated clearly to all researchers involved. 1. What is the overall objective of the analysis? 2. What are the fundamental research questions the analysis will attempt to answer? 3. 4. What are the intended outputs of the analysis? What are the required data inputs?
5. What are the required resources to conduct the analysis? 6. Who will coordinate the research? 7. What agencies with appropriate experience can review and provide constructive feedback on the research? 8. Who is the target audience for the results of the study? 9. How will the research results be disseminated?
One of the most demanding disciplines in flight test was to accustom yourself to making precise readings from the control panel in the same moment that you were pushing the outside of the envelope. This young man put his [aircraft] into the test dive and was still reading out the figures, with diligence and precision and great discipline, when he augered into the oyster flats and was burnt beyond recognitionand the [other pilots] remarked that the departed was a swell guy and a brilliant student of flying; a little too much of a student, in fact; he hadnt bothered to look out the window at the real world soon enough. Tom Wolfe, The Right Stuff Economic theory dictates the specification of econometric models of the demand for tobacco products, within pragmatic limitscompromises must be made, for example, between theoretical considerations and the availability and quality of data. In this regard, it is essential to avoid the pitfalls of data mining (i.e., avoid applying an arbitrary succession of variables to econometric analysis of data on tobacco product consumption in the hope of finding a satisfactory explanation of the latter). However, because every country or other geographical unit is different and faces a unique and distinct tobacco epidemic, detailed knowledge of the tobacco situation within a country should inform both the theoretical modeling and econometric analysis of demand for tobacco products. For example: Cross-border cigarette smuggling can play a major role in a country, in which case researchers are forewarned to control the data for its effects in order to avoid biased estimates.
A particular religion that frowns on tobacco use might be well represented in certain parts of a country, in which case this negative influence on per capita tobacco product consumption should be controlled for.
If substitutes for smoked tobacco products, such as hand-rolled cigarette tobacco or marijuana, are cheap
and easily available in a country, poorer citizens and youth may substitute these for cigarettes if the price of the latter rises significantly. This possibility should be taken into account to the extent that data availability allows. Prior research should inform the economic analysis of demand of the social, economic, and institutional characteristics of the market for tobacco products in a country. In other words, the economists and econometricians conducting the analysis should first discover and identify the characteristics of the tobacco market in their own country. Relevant prior research across all academic disciplines should also be identified to provide the necessary background information. If information is still lacking, initiate a rapid appraisal-type research study into the salient characteristics of the domestic tobacco market, consumers, tobacco companies, and so on. The advantage of detailed background research is that it helps one: evaluate the econometric results of the analysis, and judge whether the model specification and/or other parameters should be modified; and
explain the econometric results in the research report(s) submitted to colleagues, other analysts, and policy makers.
10
This chapter outlines the difficulties in choosing appropriate data variables for inclusion in demand models, and suggests ways of preparing the variables selected in order to achieve the most valid results from the analysis. While international research experience in this regard is useful, a clear and detailed understanding of ones own country situation is the best guide to the selection and preparation of data variables. Building the data set is a time-consuming part of the analysis process, as it involves the following activities: choose the variables to be included in the model of demand;
evaluate, screen, and clean the data; prepare the data for regression analysis; and conduct exploratory data analysis.
Discussion of data in this tool focuses on the most appropriate type of data for analyzing tobacco product demand, and the necessary procedures to make the data variables useable once they are selected and obtained. See Tool 2. Data for more detailed information on potential sources of data for use in analyzing economic aspects of tobacco control, including the demand for tobacco products.
the characteristics of these data series (e.g., how many observations are covered? how was the data gathered and captured? how accurate is it likely to be? how high is the incidence of missing observations?).
11
Researchers are likely to face challenges in assembling data for demand analyses. Data for a potentially significant variable, and which should therefore be included in demand model specifications, may not be available or may be of dubious accuracy. As Chaloupka, Grossman et al (1999) point out: Economists generally rely on large survey and aggregate data collected by others for different purposes that often do not contain everything that would ideally be included. In addition, the econometric analyses employing these data must attempt to control for the variety of other factors that are also likely to affect behaviour and that are varying in the real world from which these data are drawn but for which good measures are often not available.
Data Types
The type of data to assemble determines the specification of the demand model estimated, the econometric techniques used for estimation, the measurement and specification problems encountered, and ultimately the nature of the questions that can be answered by the econometric analysis. Data types relevant to demand analysis include: 1. 2.
aggregate time-series data; cross-sectional data, including: aggregate cross-sectional data, and individual-level cross-sectional data from large surveys;
3. time-series of cross-sectional data (pooled, panel, or longitudinal data). This tool focuses on the use of aggregate time-series data, in which aggregate means total (i.e., an aggregate time-series consists of a time-series of the total values of a particular variable for the particular country or region as a whole). Further, per capita measures of tobacco product consumption, income, and other variables are often used in demand analysis, and are usually based on aggregate values. For example, the annual per capita cigarette consumption of persons aged 15 and over in a country is typically calculated by dividing total annual cigarette sales by the total number of people aged 15 and over as obtained from the population census. Hence, per capita data series are discussed along with aggregate data series. There are two important reasons why researchers in certain countries may have to rely on aggregate time-series data to analyze the demand for tobacco products: In many low- and middle-income countries, individual-level data sets from large household and other surveys are not commonly available at a reasonable cost. If they are available, they may not include much data on respondents consumption of tobacco products.
12
In most low- and middle-income countries, cross-sectional data is unlikely to feature significant variance in crucial independent variables such as retail price. Most studies using cross-sectional data are undertaken in the United States, which has a fiscal federalist constitution wherein state and local governments have the power to levy their own excise taxes on tobacco products. This makes for sufficient variation in retail prices between states and cities to allow analysis of cross-sectional data.
By contrast, few low- and middle-income countries are large enough to have federal constitutions, let alone fiscal federalism. Excise taxes are typically levied by the central government at uniform rates for the whole country. Further, tobacco product manufacturers are unlikely to charge widely differing wholesale prices between sub-national geographic areas. Even if they did, transport infrastructures are usually so poor that transport costs entirely eliminate both the incentive for tobacco product wholesalers to arbitrage between areas of high and low wholesale price, and the retail price differentials. Hence, in such countries researchers must rely on the price variation inherent in time-series data. Accordingly, chapters 68 of this tool deal with the analysis of aggregate time-series data, while the Cross-Sectional Data and Pooled Time-Series and Cross-Sectional Data sections of the Specify the Demand Function chapter briefly discuss using individual-level cross-sectional data and individual-level timeseries of cross-sectional (longitudinal) data, assuming that these are available. In some countries, aggregate time-series data may not be available for a time period long enough to provide a sample of the necessary size to estimate a demand model. In this case, pooled time-series of aggregate cross-sectional data can be a useful substitute. For example, larger low- and middle-income countries may produce data on tobacco product consumption on the basis of sub-national regions (e.g., by state or province). If these data are available on a quarterly basis for a time period of several years, the pooled data sample may be large enough to allow demand estimation.
13
A per capita consumption measure is used to control for the influence of population growth on aggregate sales of tobacco products. If aggregate cigarette consumption figures are used instead, the
14
influence of population growth is factored by other estimators, particularly the time trend variable (unless the population above the threshold age is included as an extra independent variable). Data on cigarette sales are typically used as a proxy for cigarette consumption, with various attempts to control for additional consumption due to smuggling.
Arguably, it is more accurate to use average nicotine intake per person (or a proxy, such as tobacco content consumed per person) because this controls for changes in the tobacco content (and hence the nicotine content) of cigarettes over time. This in turn provides an allowance for the fact that, as the average nicotine delivery of cigarettes has fallen during the last 30 years, many long-term smokers have compensated by increasing the number of cigarettes they smoke. Such compensating behavior tends to artificially inflate cigarette sales figures, biasing downward any estimate of the impact of price and other control variables. However, data on cigarette sales are more easily available in developing countries than data on the tobacco and nicotine content of particular cigarette brands and the market shares of particular brands.
In low- and middle-income countries where consumption of bidis, kreteks, and hand-rolled tobacco is significant, the weight of tobacco consumed per capita per time period naturally provides a more accurate measure of tobacco product consumption. This assumes data are available on the quantities consumed or sold, and a plausible average weight can be used to convert such quantities into pounds or kilograms. In all instances, the definition and unit of measurement of the tobacco product quantity used in the demand analysis must be clearly stated by researchers. For example, if consumption is measured in packs of cigarettes sold, researchers must state clearly what is assumed to be the average number of cigarettes per pack. If consumption is measured in weight of tobacco consumed per capita, the assumptions and arithmetic manipulations made in arriving at this data series need to be explained clearly. To calculate per capita consumption, data are required on the size of the relevant population. Prior studies have generally assumed that significant consumption of tobacco products begins in late adolescence, thus providing one definition of a population. However, in many low- and middle-income countries, inadequate and/or poorly enforced restrictions on youths access to tobacco products means that significant prevalence of smoking and other tobacco product consumption starts at an earlier age. Regardless of the threshold age, researchers need reasonably accurate data on the size of the population above that age. In some countries it is difficult to obtain accurate data on total retail sales of cigarettes and other tobacco products. The most common way of resolving this problem is to collect official data
15
on the total excise tax revenue obtained from tobacco product sales. This data is then divided by the average tax component per pack to yield the total quantity of cigarettes sold. However, there are two potential drawbacks to this approach: Differences in retail prices between countries can result in smuggling (both large-scale, organized smuggling and small-scale, informal bootlegging) from lower-tax countries into higher-tax countries. Using official excise tax data will therefore likely overstate cigarette consumption in lower-tax countries, and underestimate it in higher-tax countries, producing upwardly-biased estimates of the impact of retail price on demand. Refer to Tool 7. Smuggling for a more detailed discussion on smuggling and methods to account for it. Data on excise tax receipts are usually compiled from wholesale rather than retail transactions (i.e., from transactions in which wholesalers buy tobacco products from manufacturers). The timing and volume of these transactions are influenced by seasonal marketing patterns of manufacturers, and introduces an artificial element of seasonality into monthly and quarterly excise tax data, and hence into tobacco product demand figures. Methods of dealing with this problem (if quarterly or monthly data are being used) by using dummy variables are discussed in the Demand Specifications for Quarterly Time-Series Data and Demand Specifications for Monthly Time-Series Data sections of the Specify the Demand Function chapter. Where possible, data on tobacco product quantities consumed and demanded should not be seasonally adjusted, since the filters used for this adjustment often distort the underlying properties of the data. This distortion is especially problematic if the data are nonstationary and cointegration techniques have to be applied (see the Apply Tests for Non-Stationarity and Cointegration, and Specify Error-Correction Models section in the Another Demand Model chapter).
Price
In many countries, the retail price of tobacco products includes either a specific or ad valorem excise tax component. (Refer to Tool 4: Design and Administration for an in-depth discussion on excise taxes.) The relationship between tax and price affects the ultimate price of tobacco products, and is therefore an important topic of consideration for tobacco control policy. However, it is not directly relevant when quantifying the impact on demand of price and other variables. This is because price, as treated for the purposes of this tool, is the ultimate value confronting the consumer, inclusive of any excise or ad valorem tax, sales tax, value-added tax, or other levy. Exceptions to this principle occur when the tax component of retail price is used either as a proxy (if retail price data are
16
unavailable), or as an instrumental variable (if retail price is found to be endogenous). In this case, the necessary adjustments to price elasticity estimates must be made, as tax elasticities of demand understate the true price elasticities to the extent that tax is a proportion of retail price. For example, in estimating the impact of tobacco excise taxes on demand for tobacco products in Papua New Guinea, Chapman and Richardson (1990) used data on the excise taxes themselves, rather than on price. Data on price for cigarettes and other tobacco products were not available, nor was information available on the relationship between excise taxes and price. If sufficient data are available, retail tobacco product prices should be averaged over product sub-types (e.g., filter, plain cigarettes) and types of sales transaction (e.g., retail single pack, carton, vending machine), weighted by estimated market share in each case. There is likely to be substantial variation between the prices of product sub-types and the prices obtained in different types of transactions, and using a weighted average price provides a realistic idea of the actual retail price of tobacco products confronting prospective buyers. Deflate price data to real terms, using the local Consumer Price Index (CPI) or the most appropriate available proxy.
Income
Deflate data on aggregate personal disposable income to real terms, using the CPI or the most appropriate available proxy. Data on population size over the threshold age should be used to calculate per capita personal disposable income values.
Expenditures on promotional activities such as promotional allowances to retailers; point-of-purchase materials; direct mail advertising; distribution of free samples, coupons, and specialty or novelty items; multiple pack promotions and retail value-added offers; endorsements; sponsorship of cultural, sporting, and other entertainment events; and sponsorship of community and other organizations.
This can prove a challenging task, particularly in countries where advertising and promotional expenditure data are not compiled by a central industry association or market research company. In
17
addition, the measure of advertising used as an explanatory variable should be the ratio of advertising on the particular tobacco product to total advertising, in order to measure the weight of the former relative to the latter. This means that advertising data should also be collected on total expenditure in each of the above media. In practice, however, such a rigorous approach is not practical if comprehensive and systematic data are difficult to obtain. Innovative approaches to this problem include that of Hu, Sung, and Keeler (1995), who used total pages of cigarette advertising in issues of Life magazine distributed in California as a representative sample of the tobacco industry media presence in that state. Advertising was modeled as an accumulated stock of cigarette advertisements per magazine issue for all previous quarters, depreciated by five percent per quarter, with a onequarter lag between advertisements and initial impact. See the Advertising and Promotional Activity subsection in the Review and Understand the Results chapter for a further discussion of incorporating such data into a model.
Smoking Restrictions
Researchers need to judge the average intensity of smoking restrictions applied in a country in order to develop an index of average smoking restrictiveness for the country as a whole, for each of the time periods under analysis. To produce this index, first calculate the intensity of smoking restrictions in each locality or province. (As a more feasible alternative, a representative sample of localities and/or provinces can be surveyed.) Measure the intensity of smoking restrictions with a numerical scale running from 0 (least intensive) to 1 (most intensive), in line with the methodology used by Wasserman et al (1991), as follows:
Score 0.00 0.25 Intensity of Smoking Restriction No smoking restrictions in place. Smoking restricted in one to three types of public place other
18
Score
No restrictions on smoking in restaurants, but smoking restricted in at least four other types of public place. Smoking restricted in restaurants, but not private workplaces. Smoking restricted in private workplaces.
If possible, weight the index of intensity for each locality by the proportion of the total population it contains. Then calculate the average index of intensity for the whole country for the year in question. If the necessary data are not available for the computation of an index of intensity of any imposed smoking restrictions, use a dummy variable to control for their introduction.
19
20
themselves the advisability of simultaneous equation versus other ways of modeling demand for and supply of tobacco products, depending on their specific circumstances.
21
Conventional demand models, which are static (i.e., they examine the impact of explanatory variables on demand only within a single time period).
Addictive demand models, which are dynamic (i.e., demand in a given time period is affected by demand in past or future periods, as well as by other explanatory variables operating within the current time period). Addictive demand models are further subdivided into models of myopic addiction and rational addiction.
In essence, addictive demand models contain the same independent variables as conventional demand models, but with the addition of future demand and/or past demand. Typical specifications of each of these three basic demand models are detailed in the Demand Specifications for Annual Time-Series Data section of the Specify the Demand Function chapter. The rationale underlying the myopic and rational addictive models is discussed in the Nicotine Addiction and the Role of Past and Future Demand subsection of the Review and Understand the Results chapter.
Semi-log, which can be either log-lin, where the dependent variable is transformed into logarithmic values of the original data observations, while the independent variables are left as levels. lin-log, where the dependent variable observations are left as levels, while the independent variables are transformed into logarithmic values.
Double-log (also known as log-log or log-linear), where logarithms are taken of both dependent and independent variables.
Theory offers very little practical guidance about the choice of functional form for demand model specification. One minor advantage of the double-log functional form over the linear form is that the estimated coefficients on price and income are, in effect, the price and income elasticities of demand. However, when using a linear specification, the elasticities are easily calculated from the values of the coefficients and data observations using the simple formulae provided in the Calculate Elasticities of Demand section of the Review and Understand the Results chapter. On the other hand, there are two drawbacks of the double-log model: It implies constant elasticities, which may not be a valid assumption for time-series data.
22
Use of the double-log functional form can yield illogical results if the demand estimations are also being used to determine the optimal level of excise tax. Since the excise tax affects the size of consumers surplus, the size of the pre-tax surplus is a key determinant of the potential tax yield of tobacco. A double-log specification of the demand function implies an infinite consumer surplus when the (absolute value of) price elasticity of demand is less than one. This in turn suggests that the product yields more tax revenue than a linear specification, and implies that tax revenue is a positive function of the tax rate. Consequently, do not choose double-log demand specifications if the results are going to be used to determine the optimal excise tax for tobacco products (Refer to Tool 4: Design and Administration for further discussion on excise taxes).
The semi-log functional forms are useful when preliminary examination of the data suggests that the dependent variable fits a logarithmic form, while the independent variables follow a linear form (or vice versa), helping to ensure a better fit of the demand specification. The semi-log form also has the advantage of not imposing the assumption of constant demand elasticities on the demand model implicit in the double-log specification. It should be noted that the lin-log specification, like the double functional form, implies an infinite consumer surplus when the (absolute value of) price elasticity of demand is less than one. This can, therefore, yield illogical results when estimates are used in analysis of optimum tax revenue. The log-lin functional form does not incur this problem, as it implies a finite consumer surplus under the same conditions of price elasticity of demand. The linear, semi-log, and double-log functional forms can be used in separate specifications, and the results compared. Methods of determining which functional form is the more appropriate are discussed in the Administer Specification and Diagnostic Tests section of the Specify the Demand Function chapter.
23
This chapter contains typical demand specifications for aggregate time-series data, using both conventional and addictive demand models and linear and double-log functional forms. Further discussion on the rationale for including each of the specified independent variables, and the regression results to be expected, is presented in the Expect Results for Quantitative Independent Variables and the Expect Results for Qualitative Independent Variables sections of the Review and Understand the Results chapter. To make comparison easier, Table 3.1 includes examples of the demand specifications discussed.
There is likely to be a high degree of collinearity between independent variables in aggregate time-series data. The relevant rule of thumbKleins Rulestates that if the correlation between independent variables A and B is greater than the correlation between A (and/or B) and the dependent variable, then either A or B should be excluded from the model.
Dependent Variable
For the reasons discussed in the Prepare the Data Variables subsection of the Build the Data Set chapter, in all examples of model specification in this chapter it is assumed that the dependent variable is cigarette consumption per potential smoker within the relevant time period (month, quarter, or year). (Note that researchers must also define precisely the unit of
24
Economic Analysis of Tobacco Demand Table 3.1 Examples of Demand Model Specifications by Data Period, Type of Demand Model and Functional Form
Data Period/Model Type/ Functional Form Demand Specification Equation Number
Annual/Conventional Linear Double-Log Log-Lin Lin-Log Annual/Myopic Addiction Linear Double-Log Log-Lin Lin-Log Annual/Rational Addiction Linear Double-Log Log-Lin Lin-Log Quarterly/Conventional Linear Quarterly/Myopic Addiction Linear Quarterly/Rational Addiction Linear Qt = b0 + b1Pt + b2Yt + b3Tt + b4SRt + b5Dm + b6Qt1 + b7Qt+1 + b8Dq2 + b9Dq4 + b10ADt + t 3.15 Qt = b0 + b1Pt + b2Yt + b3Tt + b4SRt + b5Dm + b6Qt1 + b7Dq2 + b8Dq4 + b9ADt + t 3.14 Qt = b0 + b1Pt + b2Yt + b3Tt + b4SRt + b5Dm + b6Dq2 + b7Dq4 + b8ADt + t 3.13 Qt = b0 + b1Pt + b2Yt + b3Tt + b4SRt + b5Dn + b6Qt1 + b7Qt+1 +
t
Qt = b0 + b1Pt + b2Yt + b3Tt + b4SRt + b5Dm + lnQt = b0 + b1Pt + b2Yt + b3Tt + b4SRt + b5Dm +
t t
3.10
t t
lnQt = b0 + b1lnPt + b2lnYt + b3Tt + b4SRt + b5Dm + b6lnQt1 + lnQt = b0 + b1Pt + b2Yt + b3Tt + b4SRt + b5Dm + b6lnQt1 +
t
3.11
lnQt = b0 + b1lnPt + b2lnYt + b3Tt + b4SRt + b5Dm + b6lnQt1 + b7lnQt+1 + t lnQt = b0 + b1Pt + b2Yt + b3Tt + b4SRt + b5Dm + b6lnQt1 + b7lnQt+1 + Qt = b0 + b1lnPt + b2lnYt + b3Tt + b4SRt + b5Dm + b6Qt1 + b7Qt+1 +
t t
Double-log and semi-log versions of Equations 3.133.15 have been omitted to avoid repetition. Data variables are as follows: Qt = per capita consumption of cigarettes per adult in time period t; Qt1 = per capita consumption of cigarettes per adult in time period t1; Qt+1 = per capita consumption of cigarettes per adult in time period t+1; Pt = weighted average real retail price per cigarette in time period t; Yt = real personal disposable income per adult in time period t; Tt = time trend variable in time period t; SRt = index of smoking restrictions in time period t; Dm = an intercept dummy for smoking or health information campaign in time period m; 0 prior to time period m, 1 from time period m onwards; Dq2 = an intercept dummy for expected seasonally depressed buying during second quarter of the year; 1 for second quarter, 0 otherwise; Dq4 = an intercept dummy for expected seasonal peak buying during fourth quarter of the year; 1 for fourth quarter observations, 0 otherwise; ADt = aggregate expenditure on tobacco product advertising and promotion as a proportion of all advertising expenditure in
measurement of cigarette consumption (e.g., packs of 20) and the threshold age at which persons are considered to be potential
25
[3.1]
where: Qt = per capita consumption of cigarettes per adult in year t Pt = weighted average real retail price per cigarette in year t Yt = real personal disposable income per adult in year t Tt = time trend variable in year t SRt = index of smoking restrictions in year t Dm = an intercept dummy for the introduction of an intensive smoking or health information campaign in year m; 0 prior to year m, 1 from year m onwards t = error term
[3.2]
where: Qt1 = per capita consumption of cigarettes per adult in year t1 all other variables the same as for Equation 3.1
The rationale underlying the myopic addiction model is covered in the discussion of the relevant independent variable, past demand, in the context of nicotine addiction in the Nicotine Addiction and the Role of Past and Future Demand subsection of the Review and Understand the Results chapter. 2 The thinking underlying the rational addiction model is covered in the discussion of the relevant independent variables, past and future demand, in the context of nicotine addiction in the Nicotine Addiction and the Role of Past and Future Demand subsection of the Review and Understand the Results chapter. 26
Qt = b0 + b1Pt + b2Yt + b3Tt + b4SRt + b5Dn + b6Qt1 + b7Qt+1 + t [3.3] where: Qt1 = per capita consumption of cigarettes per adult in year t1 Qt+1 = per capita consumption of cigarettes per adult in year t+1 all other variables the same as for Equation 3.1
[3.4]
lnQt = b0 + b1lnPt + b2lnYt + b3Tt + b4SRt + b5Dm + b6lnQt1 + t [3.5] lnQt = b0 + b1lnPt + b2lnYt + b3Tt + b4SRt + b5Dm + b6lnQt1 + b7lnQt+1 + t [3.6]
lnQt = b0 + b1Pt + b2Yt + b3Tt + b4SRt + b5Dm + b6lnQt1 + lnQt = b0 + b1Pt + b2Yt + b3Tt + b4SRt + b5Dm + b6lnQt1 + b7lnQt+1 + t
The following log-lin demand specification was applied to annual aggregate time-series data from China for the period 19801996: lnQt = b0 + b1Pt + b2Yt + b3Tt + b4SRt + b5Dm +
t
where: Qt = aggregate annual cigarette consumption in year t Pt = nominal retail price per cigarette in year t Yt = nominal personal disposable income per adult in year t Tt = time trend variable in year t t = error term Estimated results for this model are provided in Table 3.2 in the Administer Specification and Diagnostic Tests section of this chapter.
27
Lin-Log
Lin-log functional forms of the conventional, myopic, and rational addiction demand models specified, respectively, in Equations 3.13.3 are expressed as follows: Qt = b0 + b1lnPt + b2lnYt + b3Tt + b4SRt + b5Dm +
t
[3.10]
t
Qt = b0 + b1lnPt + b2lnYt + b3Tt + b4SRt + b5Dm + b6Qt1 + Qt = b0 + b1lnPt + b2lnYt + b3Tt + b4SRt + b5Dm + b6Qt1 + b7Qt+1 + t
[3.11] [3.12]
To simplify notation throughout the remainder of this tool, demand specifications are listed only in linear functional form, except where the discussion specifically requires the use of double-log or semi-log forms.
28
Economic Analysis of Tobacco Demand Table 3.2 Results of Cochrane-Orcutt Regression on Annual Aggregate Cigarette Consumption Data for China, 19801996
Variable 1. OLS Regression Intercept Nominal cigarette price per pack (Yuan) Nominal aggregate income (Yuan) Time trend
2
Coefficient
T-value
Probability Value
2. Cochrane-Orcutt Regression Intercept Nominal cigarette price per pack (Yuan) Nominal aggregate income (Yuan) Lagged dependent variable Time trend [Rho] Prob. > F = 0.000; Adjusted R = 0.976; h statistic = 1.25
2
Dependent variable: Logarithm of annual aggregate cigarette sales in packs per capita. Sample mean price = 1.03, therefore short-run price elasticity at sample mean = 0.54. Sample mean price = 1.03, therefore short-run price elasticity at sample mean = 0.35, and long-run price elasticity = 0.66. Source: Teh-wei Hu, Zheng Zhong Mao, Economic Analysis of Tobacco and Opions for Tobacco Control: China Case Study A Report Submitted to the World Bank, 2000.
where: Qt = per capita consumption of cigarettes per adult in quarter t Pt = weighted average real retail price per cigarette in quarter t Yt = real personal disposable income per adult in quarter t Tt = time trend variable in quarter t SRt = index of smoking restrictions in quarter t Dm = an intercept dummy for the introduction of an intensive smoking or health information campaign in quarter m; 0 prior to quarter m, 1 from quarter m onwards Dq2 = an intercept dummy for expected seasonally depressed buying during the second quarter of the calendar year; 1 for second quarter, 0 otherwise Dq4 = an intercept dummy for expected seasonal peak buying during the fourth quarter of the calendar year; 1 for fourth quarter observations, 0 otherwise
29
ADt = aggregate expenditure on tobacco product advertising and promotion as a proportion of all advertising expenditure in quarter t t = error term
It is unlikely that monthly data supports the testing of addictive models of demand. For example, Keeler et al (1993) find that a one-month period of time is not long enough to measure any potential tendency towards rational addiction. Researchers must therefore either apply a conventional demand model to monthly data, and/or aggregate monthly into quarterly data.
30
Bardsley and Olekalns (1999) apply general methods of moments (GMM) with instrumental variables to annual data on real per capita consumption of cigarettes and other tobacco products. The rational addiction model of demand is used, and instrumental variables are necessary to deal with the endogeneity of past and future consumption. Further, three leads and lags of the real average price of cigarettes and other tobacco products are used as instruments, since the theory of rational addiction holds that optimal consumption in any period depends on the past history
31
Economic Analysis of Tobacco Demand Table 3.3 Regression Results for Annual Per Capita Cigarette Demand in Taiwan, 19661995
Independent Variable Model 1 (OLS) Yes 120.970** 2.044 1.982** 2.022 [0.48] 0.00014 1.314 [0.17] 0.645* 1.917 2.509 0.654 0.393 1.244 0.053 0.086 0.251 0.730 0.047 0.111 Model 2 (OLS) Yes 124.92*** 3.028 2.076** 2.441 [0.51] 0.00014* 1.706 [0.16] 0.643** 2.681 2.482 0.699 0.419 1.604 0.125 0.231 0.269 1.319 Model 3 (OLS) Yes 116.69*** 3.359 2.022** 2.472 [0.49] 0.00012** 2.457 [0.14] 0.673*** 4.009 0.447* 1.783 0.303 1.606 Model 4 (2SLS) Yes 149.16*** 2.893 2.705** 2.322 [0.65] 0.00014* 1.713 [0.16] 0.719** 2.758 2.758 0.764 0.529* 1.775 0.105 0.193 0.189 0.821 Model 5 (2SLS) Yes 156.68* 1.835 2.609* 1.740 [0.64] 0.00019 1.600 [0.22] 0.871* 2.251 0.428 1.533 0.009 0.017
Cointegration relationship Intercept Real average retail cigarette price per pack (1991 $NT) Real per capita personal disposable income (1991 $NT) Market share of low-tar brands (%) Dummy for strengthened warning labels from 1992 onward Market share of imported cigarettes (%) Female workforce participation rate (%) Past consumption (number of packs) AR(1)
Summary statistics: Adjusted R2 Standard error of regression 0.91 3.17 0.91 3.05 0.92 2.96 0.91 3.10 0.91 3.14
Diagnostic tests (p values of test statistics): DW statistic (first-order autocorrelation) Jarque-Bera (residual normality) Ljung-Box Q (autocorrelation; 2 lags) Breusch-Godfrey (autocorrelation; 2 lags) Lagrange Multiplier ARCH (2 lags) White (heteroscedasticity) 2.01 0.16 0.94 0.87 0.67 0.27 1.95 0.26 0.99 0.99 0.69 0.22 1.96 0.18 0.98 0.98 0.50 0.21 1.93 0.23 0.86 0.96 0.64 0.22 1.63 0.13 0.45 0.49 0.57 0.15
Asymptotic t-statistics are italicized. Price and income elasticities calculated at sample means are listed in square brackets. *** Significant at 1 percent level. ** Significant at 5 percent level. * Significant at 10 percent level. Source: Hsieh, Hu, and Lin (1999)
and expected future course of prices (see the Nicotine Addiction and the Role of Past and Future
32
Economic Analysis of Tobacco Demand Table 3.4 Error Correction Model of South African Aggregate Annual Cigarette Consumption, 19701998
Variable Coefficient T-value Probability Value
Dependent Variable: Aggregate Annual Cigarette Consumption (Millions of Packs of 20) (First Differences) Real aggregate personal disposable income (first differences) Real retail price of cigarettes (first differences) Dummy to neutralize data outlier value for 1982 Lagged residual from cointegration equation Adjusted R = 0.633; DW = 1.518
2
Demand subsection of the Review and Understand the Results chapter). GMM is used to deal with the autocorrelation associated with the leads and lags.
Coefficient Tests
As their name implies, coefficient tests test restrictions on the estimated coefficients, including the special cases of omitted or redundant variables. They include the following:
case with lagged variables). This test can be applied to least squares, two-stage least squares, and binary and count models, among others. If data are unavailable for a potentially significant variable, use a proxy for that variable rather than leave it out of the demand specification. Based on the background information on the tobacco product market, as instructed in the Conduct Background Research chapter, exercise judgement as to the probable relationship between the proxy and the variable it represents, and make allowance for this in the written interpretation of the regression results. (For instance, make clear that a proxy is used for a particular variable, in case policy makers incorrectly assume that the proxy variable can be used as a policy instrument for influencing demand for tobacco products.)
Residual Tests
Most econometric software packages feature residual tests for autocorrelation (serial correlation), heteroscedasticity, autoregressive conditional heteroscedasticity (ARCH), and normality. Not all these tests are applicable to every form of model specification.
34
Due partly to the addictive nature of tobacco product consumption, the aggregate demand for tobacco products is probably stable in the short-run, and the error terms of timeseries regressions is serially correlated. In addition, as explained in the Apply Tests for Non-Stationarity and Cointegration, and Specify Error-Correction Models section of the Another Demand Model chapter, time-series are often subject to nonstationarity, thus implying some form of autocorrelation. The following are tests of first-order autocorrelation. The Durbin-Watson (DW) statistic, the most commonly applied, tests for first-order autocorrelation only. Further, this tests range of possible results includes areas of indecision, and the test is invalidated by specification of lagged dependent variables. In most applications, the Ljung-Box Q-statistic and the BreuschGodfrey Lagrange Multiplier test are preferable.
Durbins h statistic is applied if a lagged dependent variable is included in the demand specification.
The following are tests of higher-order autocorrelation. The Ljung-Box Q-statistic tests the null hypothesis that there is no autocorrelation up to a specified order of lag. Use the Q-statistic in conjunction with graphs of residuals or graphical representations (correlograms) of autocorrelation and partial autocorrelation processes in the residuals. Choosing which lag to use for the test is a practical challenge: too small a lag can result in autocorrelation not being detected at higher-order lags; too large a lag can give the test low power because significant autocorrelation at one lag can be blurred by insignificant autocorrelations at other lags. Therefore, repeat the test for a range of lags. This test can be applied to least squares, two-stage least squares, and non-linear least squares regression.
The Breusch-Godfrey Lagrange Multiplier (LM) test is an alternative to the Q-statistic, and is applicable whether or not lagged dependent variables are specified. This test can be applied to least squares or two-stage least squares regression.
The best method for dealing with autocorrelation is to identify whether a misspecification error is responsible (such as an omitted variable), and modify the regression specification accordingly. Failing this, the adopted procedure depends on the nature of the relationship between the residuals. Most econometric software packages offer corrections for first-order autocorrelation based on linear regression techniques, such as the Cochrane-Orcutt, Prais-Winsten and Hildreth-Lu procedures (reference). These have drawbacks when working with models containing lagged dependent variables, or which feature higherorder autoregressive specifications. However, current
35
econometric software packages also offer techniques based on non-linear regression. An example of a Cochrane-Orcutt correction applied to annual aggregate data on cigarette consumption is detailed in Table 3.2. The results are improved, but it is also important to note how significantly the value of short-run price elasticity is altered by the procedure.
Model the heteroscedasticity and use weighted least squares regression to obtain more efficient estimates.
Residual Normality
Econometric software packages often present tests of whether the regression residuals are normally distributed, such as the JarqueBera statistic, together with histograms (bar charts) of the distribution of the residuals. The Jarque-Bera test compares the skewness (asymmetry) and kurtosis (flatness) of the distribution of the residuals with that of the standard normal distribution. Apply the test to least squares, two-stage least squares, non-linear least squares, and binary and count models, among others, but not to cointegration regressions, whose estimates are nonnormally distributed.
36
correlation between independent variables and the error terms caused by measurement error in the independent variables, simultaneity considerations, or the combination of a lagged dependent variable with autocorrelated error terms. Least squares estimators are biased and inconsistent, and inference procedures are invalid, in the presence of these specification errors. If poor scores are achieved from Ramseys RESET, carefully examine model specification and check for each of the above types of specification error by using and evaluating the following tests in combination: 1. Graph the regression residuals and check for systematic patterns. 2. Graph a Box-Cox transformation of the independent and/or dependent variables. 3. Check the value of the Durbin-Watson (D-W) or Durbin h statistic (since incorrect functional form often results in symptoms similar to those of autocorrelation). (See the subsection Testing and Correcting for Autocorrelation in this chapter for an explanation of these two tests.)
37
estimated relation over the two sub-samples. Apply the test to least squares and two-stage least squares regression. These two Chow tests can yield conflicting results, so interpret them with care.
Multicollinearity
Analysis of demand for tobacco products is problematic due to the high number of potentially significant factors determining demand. Additionally, using aggregate time-series data is a problem because of the high degree of correlation between many of these potentially significant independent variables particularly price, income, and advertising expenditure. Evaluate the correlation between independent variables to reveal the extent to which multicollinearity is present. The inclusion of highly correlated variables results in multicollinearity and unstable estimates for the parameters of interest. Consequently, estimates of the impact of price and other factors on demand are sensitive to the inclusion and exclusion of other variables. On the other hand, excluding potentially significant variables results in an omitted variables specification error, which produces biased estimates of the impact of the included independent variables. Options to deal with multicollinearity, such as imposing theoretical restrictions or using extraneous estimates, are also problematic. Though ridge regression is an increasingly popular option, it is biased at the cost of precision; and unless conducted by an experienced econometrician, ridge regression can result in seemingly precise estimates of badly biased model specifications. Hence, there are no satisfactory practical methods for dealing with multicollinearity, and imprecise estimates may be unavoidable. If this is the case, be sure to state the implications for inference tests and interpretation of confidence intervals of the estimates obtained.
38
Economic Analysis of Tobacco Demand Table 3.5 Results of Diagnostic Tests on Error Correction Model of South African Cigarette Consumption
Test Focus of Test Parameters Result Probability Value 0.2744 0.7173 0.9063 0.0107 0.9977 0.4246
Higher-order serial correlation Autoregressive conditional heteroscedasticity Heteroscedasticity and model misspecification Normality of residuals Model misspecification Structural breaks in the model
Cross-Sectional Data
It is much easier to obtain aggregate time-series data than to collect individual-level cross-sectional data. On the other hand, there are also readily available cross-sectional data collected by national governments and the World Bank. (See the Choose the Variables section in the Build the Data Set chapter for further discussion.) For example, researchers in Bulgaria, China, Estonia, Indonesia, Vietnam and the United States have access to household consumption surveys that include data on smoking status by individual or by household. Analysis of cross-sectional data can be used as a substitution supplement for analysis of aggregate time-series data. However, use of cross-sectional data requires considerable care. Some of the most important reasons are: Information on the prices of tobacco products purchased by respondents is generally not included in household survey responses.
In most countries, cross-sectional price data does not yield sufficient variation to allow analysis of the price-responsiveness of demand.
It is not possible to distinguish the impacts of price and policy interventions from other underlying long-run determinants of demand for tobacco products by using cross-sectional data.
While aggregate data cover the consumption habits of the entire country concerned, cross-sectional survey data cover a sample of the population. The results of analysis of the data are representative of tobacco product consumption habits in the country as a whole only to the extent that the survey sample is representative of the entire population.
39
However, cross-sectional survey data representative of the general population, or of particular target groups within the general population, can provide extremely useful supplementary evidence on critically important questions regarding tobacco product consumption, as discussed below.
40
population or of particular population sub-groups who consume tobacco products). 2. The decision of how much of the tobacco product to consume daily (also known as a conditional demand). In effect, this means deciding on the desired daily intake of nicotine (and tar and other harmful substances in the case of cigarettes). This issue is related to the individuals concerns about whether or not to reduce tobacco product consumption for health reasons. 3. The decision on what sub-category and brand of tobacco product to consume (e.g., plain vs. filter cigarettes, lower-tar brands vs. higher-tar brands of cigarette). The nicotine and other content of preferred sub-types and brands influence the decision regarding desired daily consumption. 4. The decision to attempt to stop consuming tobacco products, usually for health reasons but possibly also for reasons of personal finances, social disapproval, and so forth. These decisions are approached in different ways depending on the demographic characteristics of the individual (e.g., age, gender, education level, ethnicity, community values, and income level). The nature of these processes is thus of interest to policy makers keen on extending the benefits of tobacco control to particular population sub-groups, such as youth and the poor. Individual-level data is useful for investigating issues such as the comparative price-sensitivity of various population sub-groups, especially youth and young adults. (It is of particular interest to understand how the decisions of young people to consume tobacco products are influenced, since most people begin tobacco product consumption early in life; and it is more difficult to stop the younger one starts.)
41
Price
Problem of Determining the Price Variable Finding a meaningful and statistically acceptable price variable in cross-sectional survey data is a major challenge. Most surveys do not ask for the price of a pack of cigarettes paid by the respondent. If they do report price information, only smokers respond with this information in two possible forms: total paid expenditures on their cigarette consumption or the price per pack of their chosen brand. Estimate average price per pack by dividing total expenditure by the number of packs of cigarettes consumed, and check this against independent information on the actual prices. This estimated price variable is endogenous, since the denominator of this independent variable is created from the dependent variable (i.e., quantity of cigarettes consumed). Importance of the Price Variable Price (of which tax usually comprises a large part) influences smoking behavior. An increase in price affects a smokers decision about the number of cigarettes to smoke, whether to switch to cheaper brands, or to quit smoking altogether. Price also affects the non-smokers decision of whether or not to start smoking. To understand how price influences smoking decisions, its necessary to estimate the price elasticity of cigarettes. The price elasticity measures individuals sensitivity to price changes. The price elasticity of demand for cigarettes has very strong policy implications. Once the price elasticity is known, one can determine how much to increase price in order to achieve a planned reduction in consumption, and what will be the increase in government revenue as a result of the price (tax) increase. Problem of Using Price Variable in Cross-Sectional Data Surveys collect self-reported cigarette prices from respondents who smoke. As a result, the price variable in the data can reflect endogenous choices of cigarette brands and quality. In other words, individuals exercise some choice over the price they pay for cigarettes, rather than the price they pay being determined exogenously, or entirely independently of their decisions about whether and how much to smoke. So the price variable can be correlated with unobservable differences in preferences, yielding biased estimates of the price elasticity. Because of this possibility, the price variable may be endogenous. Price Value Unassigned for Non-Smokers Whether or not they smoke, individuals face a price value for cigarettes. That is, price plays a significant role in deciding whether or not to smokefor smokers and non-smokers. In a regression, if no other price is assigned for non-smokers so that the price variable has the value zero for non-smokers, the price elasticity estimate can be positive because of the weight of the zero value in the price distribution from all the non-smokers.
42
Assign a Price Variable Value There are two major problems for finding the price variable in a cross-sectional survey: zero price for non-smokers and endogeneity of price variable. There are two approaches to resolving these problems. The first is to find published average cigarette price data collected by the government or cigarette industry in various locations, such as cities, counties, and provinces across the country. Based on the individual respondents location, the price of cigarettes can be collated to each respondent, regardless of whether the respondent is a smoker or non-smoker. This approach assumes that there is enough of a price variation across the country due to differences in local tax rates, transportation costs, income differences, and cost of living differences. On the other hand, this is a desirable approach, because it assumes that the individual consumer is a price taker and faces a market price. No endogeneity problem exists in this approach. Also, regardless of whether the individual is a smoker or non-smoker, there is a market price for cigarettes. Studies carried out by Wasserman and Hu provide examples of this approach. The second approach is to assign a price variable value for nonsmokers by estimating a regression with price as the dependent variable, and the independent variables are the individuals characteristicssuch as gender, income, and educationthat are associated with smoking behavior. It is possible to use this regression to assign a predicted price for non-smokers with specific characteristics, based on the price paid by smokers with matched characteristics (living in the same neighborhood, with similar income, same age, sex, education, occupation, family structure, etc).
43
coefficient of the gender represents males and should be interpreted by referencing females. Education The education variable can be defined in several different ways: educated versus not educated, high school and above education versus less than high school education, or several education variables such as primary, secondary, high school, technical, and so onin which case the non-educated group is provided as a reference. The education variable can have a high correlation with the income variable, creating multicollinearity. That is, as an individuals education level increases, income is more likely to increase as well. But education can also play a negative role in a smoking decision. Since more educated individuals are likely to expose the adverse health impacts of cigarette smoking, they may therefore reduce their consumption, even though their income increases. Religion Some religions outlaw smoking and require followers not to smoke. This is the case in the United States among Mormons, and in Egypt among Muslims. In countries where people have several different religious beliefs, it is necessary to create a religion variable similar to the age and gender variables. Tobacco Control Variables If tobacco control policies do not differ between states or provinces and the data is available only at one point of time, then there is no variation to include a variable for tobacco control policies in the estimation. But if a country has, for example, two years of data collected before and after major tobacco control policies take place, it is useful to compare the pre- and postpolicy impact on cigarette consumption. In this case, merge two data sets and measure the impact by creating a discrete variable. For example, if the first survey is conducted in 1996, assign those observations the value of 0, and assign the later survey observationsassuming they are taken in 1999the value of 1. The caveat for this definition is that the variable can capture the impact of other developments taking place simultaneously with tobacco control policies. To avoid a dichotomous tobacco control variable from capturing other factors, one solution is to control provinces and regions in the estimation, depending on the availability of information and the number of observations. If provinces or states implement different tobacco control policies, try to create a tobacco control variable. For example, if tobacco use is restricted only in public places, assign the variable a value of 1; if restricted all public and health places, assign the variable a value of 2; if restricted all public places, health institutions, and schools, assign the variable a value of 3; if smoking is restricted everywhere and some restrictions are brought for advertising in print and electronic media, assign the variable a value of 4; and so on. This definition works well when the time series data is more than two years. It is also possible to
44
create different tobacco control variables based on the strength of tobacco control policies applied in different provinces or states.
[( / q )] /( / t )
where: = the estimated coefficient of tax variable in the regression equation above
= the sample mean of the cigarette price
p/t = the change in cigarette prices resulting from a change in excise taxes. This can be estimated by regressing price as a function of tax: Price = + Tax + and the estimated coefficient of tax ( ) is p/t. Use Tax as an Instrumental Variable In order to estimate a demand equation, it is necessary to have a price/tax variable with many values, not just a single value for all consumers at the time of the survey. (If price/tax has only one single value for all observations, then one can only estimate a single point and not a whole demand curve.) Cigarette demand
45
studies typically get variations in price due to tax differences over time and/or across different tax jurisdictions (e.g., the 50 U.S. states all have different cigarette taxes, so even if a national survey is limited to a single year, it has considerable tax variation). However, in most developing countries household or individual surveys are cross-sectional at one point in time, with no local or provincial taxes. Further, one or two years of household data do not provide enough variation in price/taxes to include tax as a variable in the analysis. Although tax rates may not vary by provinces or states in most countries, cigarettes are often taxed at different rates based on length, production size, quality, type, manufacture process (handmade vs. machine-made), and origin. When the characteristics of cigarettes that individuals smoke are identified from the survey even one done at a single point in timethere may be enough tax variation. On the other hand, if there is no characteristic information other than price, then it is necessary to find other sources showing detailed price information by type, size, quality, origin, and so on. This information can be obtained from commerce departments and/or customs and tax administration departments of the Ministry of Finance. With this data, use price to determine the types of cigarettes smoked and assign a corresponding tax level. However, be careful about price variations of each brand of cigarettes in urban and rural areas, in different provinces or states, or even in different kinds of sales points (e.g., supermarkets, street vendors, convenience stores, and gas stores). There is no rule of thumb about how much variation is needed in a variable to get a reasonable estimate. But where taxes vary depending on several different characteristics, as mentioned above, there is usually enough variation in the tax variable to get a reasonably good estimate.
46
where: lnQjl = logarithm of average daily consumption of cigarettes by the jth individual in the lth locality Pl = real average price of cigarettes per pack in the lth locality Yjl = real disposable personal income of the jth individual in the lth locality Vjl = a vector of individual demographic characteristics such as age, gender, ethnic group, education level, marital status, and religious observance Fjl = a vector of family characteristics including family size, identity of household head (father or mother), and education level of household head SRl = index of smoking restrictions in the lth location t = error term
fr m h r , o ta o ee b in
Step A2:
Estimate the price elasticity from this formula, as such: price elasticity =
m ean predicted price consum ptio n predicted price m ean consum ptio n
= 1
To estimate tax elasticity from the model, conduct the following: tax elasticity =
consum ptio tax n m ean tax m ean consum ptio n
47
= 1 1
However, if logs are involved, estimate the price and tax elasticity differently. For example, suppose: Step B1:
lg o
( pic ) = 0 r e
+ 1 lg o
(ta ) + th r x o es
lg o
o f m r
hr , o t i ee ban
t e " pe i t d h r dce
( pic r e
1
mean consumptio n
mean tax consumptio n log ( price ) log ( price ) tax mean consumptio n
= 1
1
mean tax
= 1 1 mean consumptio n In both equations the log dependent variable of log (consumption) variable is not used because, for non-smokers, the consumption variable is zero. Since the log of zero is not defined, the observations for non-smokers are dropped from the estimation if log (consumption) is used.
48
Choose Probit or Logit at the First Step Estimate the price elasticity of decision to become a smoker from the first step estimation of probit and logit models. Since the second step estimation uses the OLS technique, the price elasticity of cigarette demand estimation depends on which model (e.g., linear, log linear, double log linear) is used. Why Probit or Logit? Probit and logit models estimate the probability of being a smoker for individual level data (or having a smoker in a household for household level data), and measure this likelihood after controlling the relevant variables used in the demand model. The dependent variable in the first step is defined as a dichotomous variable with the values 1 for smokers, and 0 for non-smokers. Apply this to the aforementioned formula to produce:
Pr ( y =1) = f ( x ' s and ' s )
where: xs = variables s = the variables coefficients In practice, the probability function is rarely a linear form, since the linear probability model allows for predicted values to be outside of the range (0,1). Instead, the probability function is usually either the standard normal density, or the logistic distribution, both of which provide predicted values within the range (0,1). If the former is used, the result is a probit model. If the latter is used, the result is a logit model. The Probit Model Estimate price elasticity of decision to smoke from the first step. The probit model appears in Equation 3.16 by using the first two specifications for the distribution of being a smoker or y = 1. Pr ( y = 1) = ( ) = x
x
1 exp z 2 dz 2 2 1
Pr ( y = 0) = 1 ( x )
( y x ) = 0 (1 ( ) ) + 1 ( ) = ( ) x x x
[3.16]
where: x = k 1 vector of independent variables = k 1 vector of coefficients corresponding to the independent variables = the standard normal cumulative distribution function Use the last equation to calculate the average change in E(y|x) with respect to the kth (price) variable:
( y x ) x k =
k 1 2 exp ( ) x 2 2
49
Note that in the probit model, the derivative of E(y|x) with respect to xk varies with the level of xk and the other variables in the model. Therefore, evaluate the derivatives at the mean values of all the x-variables in the sample. Then, determine the elasticity (at the means) of E(y|x) with respect to the kth variable with the following formula:
( y x ) x k xk k 1 = exp x ( y x ) 2 2
xk ( y x )
where: k = the coefficient on the kth (price) variable E(y|x) = the average value of the y-variable For example, suppose yi assumes value 1 if person i is a smoker, and the value 0 if person i is a non-smoker. Then, E(y|x) is the percentage of smokers in the sample, x k is the average value of the kth (price), and x is equal to:
0 x 0 + 1 x1 + 1 x 2 + k x k
where the bars represent sample averages (or mode) of the underlying variables such as income, household size, age, sex, and so on. The Logit Model The logit model appears in the third regression below, by using the first two specifications for the distribution of y.
Pr ( y = 1) = ( x ) =
exp( x ) 1 + exp( x )
Pr ( y = 0) = 1 ( x )
( y x ) = 0 (1 ( ) ) + 1 ( ) = ( ) x x x
where: x = the vector of independent variables = the vector of coefficients corresponding to the independent variables = the logistic distribution function Calculate the average change in E(y|x) with respect to the kth variable as:
( y x ) x k
= k
= k ( y x ) (1 ( y x ) )
exp ( ) x = k ( ) (1 ( ) ) x x (1 + exp ( x ) ) 2
Evaluate the derivative at the mean values of x-variables in the sample. Determine the elasticity (at the means) of E(y|x) with respect to the kth variable with the usual formula:
( y x ) x k xk xk = k ( y x )(1 ( y x ) ) k x k (1 ( y x )) ( y x ) ( y x )
50
For example, suppose yi takes on value 1 if person i is a smoker, and the value 0 if person i is a non-smoker. Then, E(y|x) is the percentage of smokers in the sample, x k is the average value of the kth variable, and k is the coefficient on the kth variable in the regression model. Estimate total price elasticity from the two-part model specified above. In the two-part model, in order to estimate overall price elasticity, first estimate probability of decision to smoke from the first step estimation as explained above for both probit and logit specification. Then estimate price elasticity of cigarette consumption conditional on being a smoker by using one of the OLS estimation techniques. Then determine the overall price elasticity by summing together the price elasticity of participation (from the first regression) and the price elasticity of consumption (from the second regression).
Price
There are likely to be several difficulties associated with the use of the price variable in cross-sectional demand analyses in most countries: Data on the prices paid for tobacco products is often not gathered in household surveys. Nevertheless, after some investigation, average prices for tobacco products prevailing in particular locations can usually be collated with the data entries of particular respondents.
Another way of dealing with lack of survey data on tobacco product prices is to calculate average prices from survey responses on expenditures and tobacco consumption. One drawback to this approach is that selfreported underestimation of tobacco product consumption can result in systematic overestimation of average prices. Another problem is that this method makes price endogenous (which is not otherwise a problem with individual-level data), but instrumental variable techniques can deal with this difficulty.
In any case, most countries do not experience significant variation in tobacco product prices between localities. In these circumstances, price is unlikely to be a
51
statistically significant variablein contrast to the results of demand analysis of aggregate time-series data. It is important to carefully explain to intended target audiences the reasons for these apparently contradictory results, in order to prevent confusion and avoid conveying the impression that price does not have an impact on tobacco product consumption.
Income
Cross-sectional data from individual or household surveys are likely to have wider variation in the income variable than aggregate time-series data, so the income variable is probably more statistically significant than in time-series studies.
Smoking Restrictions
Econometric analysis in the United States tends to find that workplace smoking restrictions both lower the daily cigarette consumption rate of current smokers, and are associated with lower incidence of smoking initiation among non-smokers. Nonetheless, it is important to keep in mind the potential simultaneity between low smoking participation rates and smoking restrictions, and to apply appropriate methods to correct for this.
52
Demographic Factors
The demographic structure (e.g., age, education level, ethnicity, and community values) of a population can be important in explaining tobacco product consumption. For example, Bardsley and Olekalns (1999) point out the importance that a declining number of older Australian smokerswho became addicted to nicotine during their military service in World War IIhas on tobacco product consumption. The influence of demographic factors is summarized as follows: Age It is popularly suggested that youth are more price-responsive than adults (Chaloupka and Warner, 1999), at least in the short run, for the following reasons: Due to the addictive nature of smoking, longterm adult smokers are less likely to adjust quickly to changes in price than young smokers (who have been smoking for a relatively short time) or young prospective smokers (who are considering smoking or in the early stages of experimentation).
Peer influence is likely to be more significant among youth than among adults.
The proportion of disposable income a youth spends on tobacco products is likely to be much higher than that of an adult smoker. Youth tend to have a much shorter time horizon (i.e., be more oriented towards the present) than adults.
53
The econometric evidence on the relationship between age and price-sensitivity of demand for tobacco products is mixed. Early studies in the United States conclude that overall cigarette demand by youth was up to three times as price-responsive as that of adults, while later studies find no significant differences in price responsiveness by age. The most recent U.S. studies (Chaloupka and Wechsler, 1997), based on much larger samples and more careful econometric specification, confirm the idea that youth and young adults are more responsive to cigarette price than older adults. Evidence also shows that, in the United States, there is an inverse relationship between the price elasticity of demand for smokeless tobacco products and age. Because of the lack of cross-sectional price variation in many low- and middle-income countries, perhaps it is possible that the only way to test these propositions about the relative priceresponsiveness of youth tobacco product consumption is with panel or longitudinal data. Gender Several studies (reference) in the United States find that men are less price-sensitive than women. In addition, an early application of the Becker and Murphy (1988) model of rational addiction to individual-level survey data finds that U.S. men behave more myopically and are relatively responsive to price than women. Race/Ethnicity Studies in the United States find that cigarette demand among African Americans and Hispanics is more price-sensitive than among white non-Hispanics. Similar differences are observed among black and white youth in the United States. However, expect collinearity between race and income in certain multicultural societies. For example, high correlation between socioeconomic status and race/ethnicity in the United States means that explanatory variables controlling for race probably reflect the influence of income on demand for tobacco products, and vice versa. Education Level Cigarette demand among U.S. youth is relatively less elastic for more educated or higher-income individuals. Similarly, in applying the Becker and Murphy (1988) model of rational addiction to U.S. individual-level survey data, it is found that less educated persons behave more myopically than their more educated counterparts. In addition, less-educated persons are more price-responsive than higher-educated persons. Community Values and Religion Certain religious denominations frown on the consumption of tobacco products. Be aware of this and other community values when modeling tobacco consumption. (Refer to the Conduct Background Research chapter for more information on the importance of community values and religion.)
54
To what extent does increased health information result in reductions in the daily consumption of cigarettes by smokers, rather than in shifts to brands with lower tar and nicotine content?
A shortcoming of cross-sectional data in many countries is the lack of price variation between localities (see the CrossSectional Data subsection of this chapter). The availability of time-series of cross-sectional data removes this difficulty. Also, while time-series data sets and cross-sectional data sets can each have only a few observations, pooling time-series of crosssectional data often provides a sample large enough for meaningful regression analysis. The basic methodological and technical issues of time-series data, as discussed in this tool, are also applicable to pooled timeseries of cross-sectional data. The simplest model is to introduce both time-dummy variables and cross-section dummy variables, the so-called fixed effect model. However, most econometric software packages now include a range of sophisticated options for estimating equations on pooled data. These include the use of pools and systems to estimate general or more complex models using two-stage least squares or non-linear specifications, and the ability to estimate models with error-component methods. Hence analysis of pooled data is likely to be more complex than using time-series or cross-sectional data, but also more rewarding.
55
56
Economic Analysis of Tobacco Demand Table 3.6 Calculation and Characteristics of Elasticities of Demand for Time-Series Data, by Type of Demand Model, Data Term, and Functional Form
Functional Form Model/Regression/ Term Double-Log (Constant Elasticity) Linear (Variable Elasticity) Log-Lin (Variable Elasticity) Lin-Log (Variable Elasticity)
bi = estimation coefficient on relevant independent variable (e.g., price, income) Xt = value of independent variable at time t Qt = value of dependent variable (quantity demanded) at time t k = partial adjustment factor = (1 estimated coefficient on lagged dependent variable)
57
and/or Qt. It is not be sensible to calculate it at extreme values of Pt and/or Qt if they vary greatly from the mean values.3 Short-run elasticities of demand for other independent variables (e.g., income) are calculated in the same manner.
[3.20]
[3.21]
and rearranges Equation 3.20 accordingly, the demand function to be estimated becomes: Qt = kb1Pt + (1 k)Qt1 + k
t
[3.22]
Hence, calculate the price elasticity of demand4 according to the following forms.
Refer to Table 3.3 for examples of price and income elasticities of demand calculated from data on annual per capita cigarette demand in Taiwan. 4 Derived from Equation 3.2. 58
Price
The literature on price-demand focuses on how price changes influence both the decisions of whether or not to smoke and how many cigarettes to smoke (given that one chooses to smoke). The
5
See the Apply Tests for Non-Stationarity and Cointegration, and Specify Error-Correction Models section of the Another Demand Model chapter. 59
responsiveness of demand to price changes is measured by the price elasticity of demand, which is defined as the percentage change in demand resulting from a one percent change in price. Economic theory predicts that demand and price changes move in opposite directions (i.e., if price rises, demand falls), so the numerical measure of price elasticity of demand is expected to be negative. For example, a 0.5 value for price elasticity of demand means that a 1 percent rise in price results in a 0.5 percent decline in demand. Most recent conventional models of cigarette demand estimate price elasticities of consumption ranging from 0.14 to 1.23, but most results from developed countries fall within the narrower range 0.3 to 0.5. Although there are numerous studies of the price-demand relationship in industrialized countries, studies on developing countries have occurred only in the last ten years. A small but growing number of studies examine the demands for cigarettes and other tobacco products in a few low- and middleincome countries, while new research is beginning to focus on others. Arguably, economic theory suggests that demand in low- and middle-income countries is more responsive to pricedue to relatively low incomesthan demand in wealthier countries. In general, the findings from econometric studies support this hypothesis, suggesting that cigarette demand in lower-income countries is at least twice as sensitive to price as demand in higher-income countries. Evidence on own-price elasticity of demand from low- and middle-income countries includes the following estimates: Papua New Guinea: Chapman and Richardson (1990) are the first to empirically estimate the impact of tobacco taxes on the demand for tobacco products in a developing country. The price elasticities of demand suggested by their estimated tax elasticities are 1.42 for cigarettes and 1.00 for other tobacco products.
Turkey: Tansels (1993) estimates average 0.21 for short-run and 0.37 for long-run price elasticity of demand for cigarettes. This is an exception to the findings of most recent demand analyses for low- and middle-income countries, as they are lower than expected.
China: Several recent studies produce estimates of the price elasticity of demand in China as a whole, Sichuan province, and Taiwan in ranges centering on 0.75.
South Africa: Recent studies estimate the shortrun price elasticity of demand for cigarettes at 0.59 and the long-run elasticity of demand at 0.69.
Zimbabwe: A recent study estimates a price elasticity of demand for cigarettes of 0.85.
60
Brazil: A study by Costa e Silva (1998), using limited data, produces a suggestive result for price elasticity of cigarette demand of 0.11 in the short-run and 0.80 in the long-run.
Income
Early analyses of the demand for cigarettes in industrialized countries find significant and positive income elasticities of demand, while more recent studies find insignificant and negative income elasticities. In the United States, for example, the income variable has a statistically significant positive impact on cigarette consumption, though it was negatively significant during the 1980s and 1990s. When United States data from the 1960s to the 1990s is combined, income is statistically insignificant. This suggests that tobacco products have moved, over the past 30 years, from being a superior good (or at least a normal one) to one preferred more and more by those in lowerto lower-middle income categories. This may be due to higher education levels of those in the higher income groups, which in turn brings increased awareness and appreciation of the health risks of smoking. In some low- and middle-income countries, by contrast, tobacco products (particularly manufactured cigarettes as opposed to hand-roll cigarette tobacco, bidis, and kreteks) still have the status of superior goods, and hence positive income elasticity of demand, as found in the following studies: Chapman and Richardson (1990) find strong, positive effects of income on demand for both cigarettes and other tobacco products in Papua New Guinea.
Tansel (1993) finds a strong, positive effect of income on cigarette demand in Turkey.
Studies from other low- and middle-income countries report results for income ranging from a statistically significant positive impact on tobacco product consumption to a significantly significant negative impact.
61
Promotion covers a wide variety of activities, including: promotional allowances to retailers; point-ofpurchase materials; direct mail advertising; distribution of free samples, coupons, and specialty or novelty items; multiple pack promotions and retail value-added offers; endorsements; sponsorship of cultural, sporting, and other entertainment events; and sponsorship of community and other organizations.
Tobacco product manufacturers argue that cigarette advertising has no impact on aggregate cigarette consumption, but simply affects market share of particular cigarette brands while providing useful information to smokers about brands, including their nicotine and tar contents.
Public health advocates contend cigarette advertising and promotion directly influence cigarette consumption, in that: Children and young adults are enticed to experiment with smoking and to initiate regular smoking.
Current smokers willingness to quit smoking is reduced. Smokers increase their daily cigarette consumption because advertising serves as a cue or stimulus to smoke. Former smokers are induced to resume the habit because advertising reinforces the attractions of smoking.
Further, it is argued that cigarette advertising and promotion indirectly influence smoking by:
62
discouraging a full discussion of the health consequences of cigarette smoking in media dependent on tobacco advertising
discouraging institutions dependent on tobacco industry promotional and other support from creating political opposition to strong tobacco control policies.
Failing to account for the cumulative or stock effects of advertising can lead to an omitted variables problem, although the evidence is mixed concerning the durability of cigarette advertising.
The measure of cigarette advertising used as an explanatory variable is the ratio of cigarette advertising to total advertising.
Few studies follow Schmalensees suggestions: nearly all treat cigarette advertising expenditure as exogenous, while most use absolute or per capita measures of cigarette advertising expenditure. It is therefore not surprising that results from this research are inconclusive. Include advertising as an independent variable only when using quarterly or monthly data. The mixed evidence from studies using aggregate data suggests that aggregate cigarette advertising has a small or negligible effect on aggregate cigarette sales. Apart from the methodological shortcomings discussed above, however, it has been pointed out that advertising expenditure is subject to diminishing marginal productivity, and that the advertising response function therefore takes the form of an S-shaped curve, shown in Figure 3.1. In highly concentrated cigarette markets with limited price competition, where the personality of brands created by advertising is important to
63
Economic Analysis of Tobacco Demand Figure 3.1 The Advertising Response Function
Figure 1
National level
Consumption
0
Advertising
A1
consumers, annual aggregate advertising expenditure is typically raised to a level (for example, level A1 in Figure 3.1) where the marginal impact of total advertising expenditure on cigarette demand is negligible. This is why studies using annual aggregate data on advertising expenditure generally find it to have an insignificant influence on tobacco product demand. Advertising should thus be included as an independent variable only when quarterly or monthly data are being used. Other methodological shortcomings of econometric studies of the impact of cigarette advertising on cigarette consumption include: the lack of appropriate measures of advertising exposure and other problems with the measures of advertising employed
the failure to distinguish between the impacts of advertising and promotional activities
problems with simultaneity between advertising expenditures and sales the omission of other key variables, such as health information or counter-advertising campaigns taking place at the same time concentration on restrictions in one or two advertising media, without controlling for the ability of cigarette manufacturers to shift expenditure towards other media and into promotional activity, and to develop new marketing approaches6
In the United States, for example, bans on tobacco advertising via broadcast media caused the proportion of total advertising and promotional expenditure channeled into traditional advertising media to fall from 87 percent in 1974 to 10 percent in 1996even though total advertising and promotional expenditure increased dramatically during that period. 64
failure to account for the fact that expenditure on advertising campaigns may be committed (and thus enter advertising statistics) in one time period, but be spent only during the following several time periods
More appropriate approaches include the examination of much more disaggregated data and the analysis of non-marginal changes in advertising expenditures, such as those resulting from significant restrictions or complete bans on cigarette advertising and promotion. A number of country-specific studies examine the impact on smoking of restrictions and bans on cigarette advertising. In general, results suggest that partial advertising bans lead to temporary reductions in cigarette smoking, but have little effect in the long-run, while more extensive restrictions and complete bans on cigarette advertising combined with strong health warnings on cigarette packs lead to more permanent reductions in smoking. It should be clear that there are many ways of modeling the impact of tobacco product advertising and promotional expenditure. For example, if the introduction of a health information campaign significantly counteracts the influence of tobacco product advertising and promotion, specify a slope dummy to multiply with the variable ADt in Equations 3.73.9. Depending on ones view of the most plausible way to model the influence of the health campaign in question, the slope dummy can take the value 1 prior to the start of the campaign, and some value less than 1 (whether constant, declining, or increasing to 1) thereafter.
65
Some myopic addiction models treat preferences as endogenous, allowing them to change in response to past consumption (and in some cases to other factors, such as advertising). Other myopic addiction models allow past consumption to influence current consumption via an accumulated stock of past consumption. Myopic addiction models predict that the long-run price elasticity of demand for tobacco products will be larger than the short-run price elasticity of demand (in absolute value).
Indeed, as pointed out in The Economic Case for Demand Intervention section of the Define the Objectives of the Analysis chapter, the observation of widespread information failures regarding the health consequences of smoking and the risk of addiction provide an economic rationale for government intervention in the tobacco market. 66
Hsieh and Hu (1997) estimate several alternative specifications for Taiwan, including one allowing for myopically addictive behavior. Current smoking is positively related to past consumption, consistent with myopic addiction.
Van der Merwes (1998) estimates for South Africa do not support the hypothesis of rational addiction.
In common with studies of high-income countries, studies that estimate models of addictive behavior in low- and middleincome countries provide mixed support for the hypothesis of rational addiction, but are more generally supportive of myopic addiction. This implies that the long-run reductions in cigarette smoking and other tobacco use resulting from a price increase exceed the short-run effects.
Time Trend
Aggregate demand for tobacco products varies with timeif there is no variation in demand over time, there is no point in regressing a time-series of aggregate demand on a set of independent variables. A time trend is often specified among the independent variables in previous econometric analyses of tobacco product demand. This is typically justified on the grounds that:
It is important to take into account a secular trend in tobacco product demand in the population concerned. The first justification is hardly adequate. There must be sound a priori theoretical and practical reasons for including an independent variable in the demand specification, and blindly following the examples of others is not a good idea. Regarding the second justification, note that there is actually no such thing as a secular trend in demand for tobacco products. Changes in demand for tobacco products are driven by such factors as price, income, advertising, addiction, counteradvertising, and smoking restrictions (as discussed in this and the following sections). A change in aggregate demand for tobacco products over time, and not explained by such factors, can only be driven by some other explanatory variable(s).8 It makes little sense to say that demand for tobacco products is influenced simply by the passage of timeexcept in the case of non-stationarity of the data (i.e., when a time-series shows (usually an upward) drift or deterministic trend over time). It is important to test for data non-stationarity, and if appropriate to apply cointegration regression techniques (see the Apply Tests
8
For example, change in demand can be due to a change in the demographic structure of the sample population. It is suggested that a major determinant of the decline in the prevalence of smoking in Australia during the 1970s and 1980s was the decline of a generation of male smokers who were addicted to nicotine during World War II, when cigarettes were freely supplied to the armed forces. 67
for Non-Stationarity and Cointegration, and Specify ErrorCorrection Models section of the Another Demand Model chapter), without including a time-trend variable in the demand model specification. If the variables are stationary, or non-stationary but not cointegrated, include a time trend in the initial demand specification in order to test whether this variable picks up the influence of some other factor or factors not included. If it does, identify and specify that variable in the demand specification and determine the impact this has on the estimation results. Follow this procedure: 1. Regress trend-stationary variables on time and their (stationary) residuals used as substitute time-series. Difference non-stationary variables until they are stationary. 2. Include a time trend variable in the demand specification along with the now stationary variables in order to test the influence from any omitted variables. 3. If the time trend variable is significant, specify and estimate again the demand model with the original data variables in order to capture the influence of the formerly omitted variable. In the case of non-stationary variables, be alert to the dangers of spurious regressions.
68
Warning labels on cigarette packaging and advertising. Non-economic evidence from the United States and other countries suggests that multiple, strong, and direct messages that are prominently displayed are effective in discouraging smoking. The limited econometric analysis conducted on the impact of health warning labels suggests that they have led to small but significant reductions in cigarette smoking.
Mass-media counter-advertising campaigns. These have been widely used to discourage smoking and other tobacco use. In the United States, several states have earmarked cigarette tax increases to fund health education campaigns to reduce cigarette smoking. Econometric analyses of anti-smoking publicity and paid counter-advertising in the United States and Finland, the United Kingdom, Greece, Turkey, and Australia generally conclude that such campaigns have significantly reduced cigarette smoking.
Little econometric analysis of this issue has been performed in low- and middle-income countries, although Tansel (1993) finds negative and significant effects on cigarette demand in Turkey of various indicators of health information. As with all qualitative variables, there are many ways to model the impact of the health campaign intercept dummy variable Dn included in the demand specifications given in the Specify the Demand Function chapter. For example, if the influence of the health campaign is expected to diminish rapidly after its inception, rather than have Dn consistently maintain a value of 1 from year n, follow Warners (1977) example and set 0.5(tn) as its value. This variable therefore assumes values of 1 in year n, 0.5 in year n+1, 0.25 in year n+2, and so on. Be sure to apply knowledge of conditions in specific tobacco product markets in order to model the impact of this and other qualitative variables in the most appropriate and effective way.
Smoking Restrictions
As information on the health consequences of exposure to environmental tobacco smoke (ETS)also called passive smokingbecomes widespread, governments at national, subnational, and local levels in several industrialized countries are adopting policies to limit smoking in public areas and private workplaces. Although these restrictions are intended to reduce non-smokers exposure to ETS, they also lead to significant reductions in cigarette smoking, since they reduce opportunities to smoke and thereby raise the cost or full price of smoking. Evidence suggests that this is particularly true for restrictions on smoking in the workplace. In addition, restrictions on smoking can help alter the social acceptability of smoking. Several recent econometric studies of the impact of smoking restrictions on cigarette demand in the United States and elsewhere find that restrictions on smoking in public areas and private workplaces reduce both smoking prevalence and average
69
daily cigarette consumption among smokers. However, a possible methodological complication should be noted: it is found that smoking restrictions may be endogenous with respect to cigarette consumption. For example, U.S. states with the strongest smoking restrictions (those with limits on smoking in private workplaces) are also the states in which anti-smoking sentiment is relatively high and smoking is relatively low. After controlling for this, recent econometric studies show mixed results on whether the strongest smoking restrictions significantly impact cigarette demand. Overall, however, the available evidence tends to suggest that comprehensive restrictions on smoking in public places significantly reduce smoking even after accounting for their potential endogeneity. Note that if an index of smoking restrictions, as suggested in the Prepare the Data section of the Build the Data Set chapter, cannot be compiled due to lack of data, specify a dummy variable in Equations 3.13.9 to control for the introduction of such restrictions in place of SRt, as follows: Dn = an intercept dummy for the introduction of comprehensive restrictions on smoking in public places and private workplaces in time period n; 0 prior to time period n, 1 from time period n onwards.
70
71
useful guide to the difficulties involved in econometric research on the topic. Keeler, Hu, Barnett, and Manning (1993) apply the theory of rational addiction to monthly aggregate data from California. In so doing, they provide a useful and comprehensive overview of many of the practical issues involved in timeseries analysis of demand for tobacco products. Bardsley and Olekalns (1999) institute a rigorous and thorough application of the theory of rational addiction to annual aggregate data on tobacco product consumption in Australia, supplemented by data on advertising, regulatory intervention, and demographic factors.
72
VIII. Another Demand Model: Error Correction Models and Diagnostic Tests
Definitions
Cointegration
Cointegration can be defined as follows: In economic terms, two or more time-series are cointegrated if the series move together over time and the differences between them are stable (i.e., stationary), even though each series contains a stochastic trend and is therefore non-stationary. Hence, cointegration reflects the presence of a long-run equilibrium to which an economic system converges over time. The differences (or error terms) in the cointegrating equation are interpreted as the disequilibrium error at each particular point in time.
In econometric terms, two or more time-series which are non-stationary of order I(1) are cointegrated if a linear combination of these series exists that is stationary, or I(0). The vector of coefficients creating this stationary series is the cointegrating vector.
Aggregate consumption and aggregate income are good examples of two cointegrated economic time-series. If they are not cointegrated, consumption can drift significantly above or below income in the long-run, as consumers irrationally overspend or spend an unusually low proportion of their income, respectively.
Collinearity
Collinearity refers to the existence of a linear relationship between two explanatory variables in a regression model. In other words, the coefficient of correlation between the two
73
variables will tend towards unity. If more than two explanatory variables within a regression are linearly related, the term multicollinearity is used. A high degree of collinearity or multicollinearity is potentially problematic because a linear relationship between two or more explanatory variables makes it difficult to separate out the influence each variable has on the dependent variable. The practical consequences of high multicollinearity include large variances and covariances in ordinary least squares regression estimators, wider confidence intervals for sample population parameters, seemingly insignificant t-ratios, and sensitivity of estimators to small changes in the data. Because multicollinearity is basically a sample phenomenon, and the extent thereof will vary from sample to sample, there is no unique means of testing for it. Instead, several rules of thumb can be applied to gauge the extent of multicollinearity. There is little one can do about serious multicollinearity, as the remedies available generally have their own potentially serious shortcomings. For example, dropping one of the collinear explanatory variables may lead to specification bias. However, researchers should be alert to the difficulties that a high degree of multicollinearity can impose on the interpretation of regression results.
Error-Correction Model
A long-run cointegration relationship between two or more nonstationary data series corresponds to an embedded errorcorrection mechanism, which compensates for short-run deviations of the data from the equilibrium of the long-run model. In other words, if the system deviates from equilibrium in the short-run, short-run demand relationship will intervene to correct this error and move the system back towards equilibrium. Both the Engle-Granger and Johansen procedures (need references) can be used to estimate the coefficients of this dynamic error-correction model (ECM), thereby obtaining the parameters of the short-run demand relationships present within the model.
The Johansen procedure, a systems approach, is more complex than the Engle-Granger method and is available only in some econometric software packages. The econometric literature has given this procedure a lot of attention, and it is currently regarded as the state-of
74
the-art in econometric analysis of time-seriesbut it requires careful handling. The Johansen procedure treats all variables as potentially endogenous, modeling each variable as an unrestricted vector autoregression (VAR) involving a number of lags. An advantage is that dynamic relationships among endogenous variables are modeled without strong a priori restrictions. Using a vector error-correction model (VECM), differences in these endogenous variables are expressed in terms of short-run changes (differences) and long-run changes (levels). These results are then decomposed into coefficients representing the speed of short-run adjustment to disequilibrium, and coefficients representing the long-run cointegrating relationship. Researchersshould carefully consider the following characteristics of the Johansen procedure: The Johansen procedure is extremely dataintensive, and is intended for application to long timeseries. Therefore, apply it only to quarterly or monthly data. Results from annual data sets are likely to be subject to significant small sample bias and error.
Illogical model specifications may result because the Johansen procedure tries to make the model as general as possible and its use of a VAR-type model places no prior restrictions on potentially endogenous variables. For example, the procedure may include lags for particular variables (such as income) that are not plausible explanatory variables in relation to the demand for tobacco products.
Johansens cointegration test can operate under several sets of assumptions regarding the presence or absence of deterministic trends present in the data, any of which can be specified in advance by users. Exogenous variables (such as seasonal dummies), if any, should also be specified in advance when applying the test.
Price Endogeneity
The market for tobacco products consists of the interaction of supply and demand. In a competitive market, both price and quantity will vary until the market clears (i.e. quantity supplied is equal to quantity demanded at a given price). Price and quantity influence each other (for example, if the price asked by suppliers falls, the quantity demanded by consumers increases) until the market-clearing quantity and price are reached. In such a situation, the market price of tobacco products is determined by the interaction of demand and supply within the market itself, and price is considered endogenous, or determined within. If the price of tobacco products is set outside this interaction of supply and demand (for example if prices are set by legislation completely independently of demand or supply), price is said to be exogenous, or determined outside.
75
If price is endogenous, the simultaneous determination of quantities and prices of tobacco products bought and sold in competitive markets causes an identification problemone can never be sure that two different combinations of market-clearing price and quantity both lie on the same demand curve. Failure to account for this in regression analysis of demand will result in biased estimates. However, if it is reasonable to assume that the supply of tobacco products is infinitely elastic (and hence that price is basically exogenous), there is no identification problem it is certain that every combination of price and quantity in the data lies on the demand curve. In many low- and middle-income countries it may indeed be the case that the supply of tobacco products is infinitely elastic, particularly for those countries that must import the bulk of their tobacco leaf and/or manufactured tobacco products. Most countries are small relative to the global market for tobacco leaf and tobacco products, and tobacco companies allocate their (finite) supplies of both to the countries that pay the highest prices. Once the marginal demander of tobacco is not willing to pay these prices, supplies are shipped to another country. In these circumstances, supply within many countries can be characterized as infinitely elastic. If it is not reasonable to assume that supply is infinitely elastic in other words, if price is endogenousappropriate econometric techniques, such as the use of instrumental variables or simultaneous equation modeling, should be applied. Hausmans test can be used to determine whether price is exogenous or endogenous.
Stationarity
Time-series data variables are characterized with respect to the stationarity of their underlying data generation processes. Accordingly, a particular data series can be categorized as being one of the following: Stationary: The data series has a constant mean and variance that are independent of time. In other words, the series fluctuates around its mean value within a finite range, and does not show any distinct trend over time.
Trend-stationary: The data series has a constant variance around a deterministic (i.e., fixed) time trend. The series fluctuates around the time trend within a finite range.
Non-stationary: The data series does not have a constant mean or a constant variance, but follows a stochastic (i.e., random) time trend, drifting either upwards or downwards over time, without being confined within a finite range. Variance increases with sample size.
76
Most economic time-series are not stationary, but trend upwards over time. It is important to control for such trends in regression analysis, otherwise spurious regressions can result. The results of such spurious regressions would suggest that there is a statistically significant long-term relationship between the variables in the regression model, when all that is being reflected is a correlated time trend rather than a meaningful causal relationship. It is therefore important to test time-series data for nonstationarity, and if it is present, to apply the appropriate econometric techniques (such as cointegration regression) to control for it.
and heteroscedasticity, and then apply whichever econometric techniques promise to solve the problems. Given the potentially serious impact of data non-stationarity on time-series analysis, however, a more desirable sequence is recommended, as follows: 1. Test all variables for non-stationarity (see the Non-Stationarity and the Problem of Spurious Regression subsection of this chapter). a. If two or more variables are non-stationary, test for cointegration (see the Cointegration Relationships subsection of this chapter). i. if a cointegration relationship exists, take its parameters as long-run model results, and estimate a short-term error-correction model (see the Estimate an Error-Correction Model and Short-Run Demand Relationship subsection of this chapter). ii. if a cointegration relationship does not exist, apply OLS regression, with a time-trend variable included in the initial demand specification used, being alert to spurious regression results. b. If one or no variables are non-stationary, apply OLS regression, with a time-trend variable included in the initial demand specification used. 2. Apply appropriate specification and diagnostic tests (see the Administer Specification and Diagnostic Tests section of the Specify the Demand Function chapter). 3. Analysis regression results and relevant specification and diagnostic test results, and decide whether: a. alternative regression techniques should be applied, and/or b. the functional form of the demand model should be re-specified, and/or c. other variables should be included or currently included variables omitted, and/or d. other data variables should be substituted where appropriate (assuming that accurate proxy data series are in fact available). 4. Once valid regression results have been obtained, calculate and review elasticities of demand (see the Review and Understand the Results chapter).
78
Apply Tests for Non-Stationarity and Cointegration, and Specify Error-Correction Models
A common and potentially serious problem in time-series analysis is non-stationarity, which can result in spurious regressions. Cointegration regression addresses this problem. Cointegration regression is difficult, and a comprehensive treatment of the topic cannot be provided in this tool. However, the remainder of this chapter summarizes salient points and highlights potential difficulties and issues. (Note: time trend variables are excluded from stationarity tests and cointegration regression specifications.)
Trend-stationary: The data series has a constant variance around a deterministic (i.e., fixed) time trend. The series fluctuates around the time trend within a finite range.
Non-stationary: The data series does not have a constant mean or a constant variance, but follows a stochastic (i.e., random) time trend, drifting either upwards or downwards over time, without being confined within a finite range. Variance increases with sample size.
Most economic time-series are not stationary, but trend upwards over time. It is important to control for such trends in regression analysis, otherwise spurious regressions can result. The results of such spurious regressions would suggest that there is a statistically significant long-term relationship between the variables in the regression model, when all that is being reflected is a correlated time trend rather than a meaningful causal relationship. The deterministic time trend in a trend-stationary variable is removed by regressing the variable on time. The resulting residuals form a new trend-free, stationary variable. Alternatively, the trend is controlled by including a deterministic
79
time trend in the regression equation.9 In contrast, regressing a non-stationary variable on a time trend generally does not result in a stationary variable. Instead, the variable must be differenced until it is stationary; the number of times the series must be differenced corresponds to the number of unit roots present in the data generating process underlying the time-series. That is, if a series must be differenced d times before it becomes stationary, it contains d unit roots and is said to be integrated of order d, denoted I(d). It is important to test time-series data for the presence of unit roots, and to apply the appropriate econometric techniques to control for non-stationarity of variables. In this regard, cointegration regression avoids spurious regressions, which are a major problem of conventional econometric analysis of timeseries data.
If the series does not seem to show any trend and has a non-zero mean, include only a constant.
If the series seems to be fluctuating around a zero mean, do not include both a constant and a trend. Alternatively, apply a sequential testing procedure to avoid arbitrary test results; the ADF test is performed with intercept, trend, and lagged variables included. If a variable is not significant at the 10 percent level, the least significant variable (as measured by the t-value) is removed, and the test equation is
9
Refer to the Time Trend subsection of the Review and Understand the Results chapter for a discussion of the role of time trend variables. 80
run again without it.10 Repeat this process until the remaining variables are significant at the 10 percent level. The ADF test is a useful indicator of stationarity, but regard its results with caution, since the test may not be sufficiently robust in small samples to discriminate between non-stationary and trend-stationary data generation processes.
Table 3.7 ADF Test Results for South African Cigarette Demand Data, 19701998
Variable Testing whether Stationary in: Levels First differences Levels First differences Levels Levels Levels First differences Trend Intercept Number of Lags 1 0 1 0 1 0 0 0 T-value Level of Integration I(1) I(0) I(1) I(0) I(0) I(0) I(1) I(0)
Aggregate cigarette consumption Aggregate cigarette consumption Per capita cigarette consumption Per capita cigarette consumption Real aggregate personal disposable income Real per capita disposable personal income Real retail price of cigarettes Real retail price of cigarettes
*** Significant at the 1 percent level 10 Significant at the 5 percent level ** The only exception to this procedure is the intercept: if the Source: Van Walbeek (2000) is not, the intercept is not excluded from the test equation.
81
is also stationary in levels. The real retail price of cigarettes is stationary in first differences, or I(1). Hsieh, Hu, and Lin (1999) apply the ADF test to annual timeseries data from Taiwan for the period 19661995, including a constant term and time trend in the test equation. The following variables are all I(1) in levels (and stationary in first differences):
per capita cigarette consumption per capita consumption of imported cigarettes real average retail cigarette price per pack
real average retail cigarette price per pack of domestic cigarettes real average retail cigarette price per pack of imported cigarettes the market share of imported cigarettes the participation rate of the female labor force
Interestingly, per capita consumption of domestic cigarettes and real per capita disposable income are I(0). It is highly unlikely that only one time-series in a demand model is non-stationary. However, if this is in fact the case, do not apply further cointegration regression procedures; instead use OLS and subject the results to diagnostic testing (see the Administer Specification and Diagnostic Tests section of the Specify the Demand Function chapter). If two or more variables are nonstationary, proceed with cointegration testing and the estimation of an error-correction model.
Cointegration Relationships
Even if data series are non-stationary, it is possible to infer a long-term causal relationship between them if they are cointegrated. Cointegration can be defined as follows: In econometric terms, two or more I(1) timeseries are cointegrated if a linear combination of these series exists that is I(0)that is, stationary. The vector of coefficients creating this stationary series is the cointegrating vector.
In economic terms, two or more time-series are cointegrated if the series move together over time and the differences between them are stable (i.e., stationary), even though each series contains a stochastic trend and is therefore non-stationary. Hence, cointegration reflects the presence of a long-run equilibrium to which an economic system converges over time.11 The differences (or error terms) in the cointegrating equation are interpreted as the disequilibrium error at each particular point in time.
11
In this instance, equilibrium is understood as a steady-state relationship between variables that are themselves evolving over time. 82
Aggregate consumption and aggregate income are good examples of two cointegrated economic time-series. If they are not cointegrated, consumption can drift significantly above or below income in the long-run, as consumers irrationally overspend or spend an unusually low proportion of their income, respectively. After determining which of the data variables are non-stationary (see the Test for Non-Stationarity subsection of this chapter), the next step is to test the I(1) data series for cointegration relationships.12 In effect, this involves estimating the long-run relationship, if any, between the variables concerned, as discussed in the Cointegration Testing: Estimate the Long-Run Demand Relationship subsection of this chapter. Note that if there are k potentially endogenous variables in the system being estimated, each of which has one unit root, there can be anything from a zero to k1 linearly independent cointegrating relationship. The presence of more than one cointegrating relationship in a demand system complicates the interpretation of relationships between the variables, but econometric software packages provide tests to highlight the most likely cointegration relationship.
The Johansen procedure, a systems approach, is more complex than the Engle-Granger method and is not available in all econometric software packages (refer to the Resources Required subsection of the Define the Objectives of the Analysis chapter for recommended software packages). The academic literature gives this procedure a lot of attention, and it is currently regarded as the state-of-the-art in econometric analysis of timeseriesbut it requires careful consideration.
The presence of I(2) variables, fortunately uncommon, makes cointegration estimation more difficult. The most recent versions of some popular econometric software packages can now handle this, however. 83
OLS estimators are super-consistent (i.e., due to the presence of cointegration they converge to their true values at a faster rate than conventional OLS estimators with stationary variables).
Johansen Procedure
The Johansen procedure treats all variables as potentially endogenous, modeling each variable as an unrestricted vector autoregression (VAR) involving a number of lags. An advantage is that dynamic relationships among endogenous variables are modeled without strong a priori restrictions. Using a vector error-correction model (VECM), differences in these endogenous variables are expressed in terms of short-run changes (differences) and long-run changes (levels). These results are then decomposed into coefficients representing the speed of short-run adjustment to disequilibrium, and coefficients representing the long-run cointegrating relationship. Carefully consider the following characteristics: The Johansen procedure is extremely dataintensive, and is intended for application to long timeseries. Therefore, apply it only to quarterly or monthly data. Results from annual data sets are likely to be subject to significant small sample bias and error.
Illogical model specifications may result because the Johansen procedure tries to make the model as general as possible and its use of a VAR-type model places no prior restrictions on potentially endogenous variables. For example, the procedure may include lags for particular variables (such as income) that are not plausible explanatory variables in relation to the demand for tobacco products.
Johansens cointegration test can operate under several sets of assumptions regarding the presence or absence of deterministic trends present in the data, any of which can be specified in advance by users. Specify exogenous variables (such as seasonal dummies), if any, when applying the test. Wherever practical, begin with the Johansen multivariate procedure rather than the Engle-Granger single-equation approach (except in the very unlikely case that only two variables are involved). Starting with a systems approach prevents consideration of only one cointegrating relationship between the variables, when in fact there can be more. Not allowing for the possibility of other cointegrating vectors results in inconsistent and inefficient estimates.13 Use the single-equation approach only when there is certainty of a single, unique cointegration vector, and when all the independent variables are known to be weakly exogenous. As this knowledge is seldom available in advance without thorough testing, it is generally safer to begin with the Johansen procedure wherever possible.
13
The presence of more than one cointegrating relationship between variables can cause difficulty in practical economic interpretation of the causal relationship between variables. 84
A practical example of the difficulties involved in applying cointegration analysis is provided in Table 3.8. Both the Johansen and Engle-Granger methods are applied to test cointegration, but only the Engle-Granger results are published. While the overall superiority of the Johansen method is fully appreciated, the Engle-Granger results are relied upon instead for the following reasons: Data for all relevant variables are available only on an annual basis, giving a sample too small for the Johansen procedure to yield reliable results.
A priori restrictions are placed on the specification of the demand model, and those restrictions are ignored by the Johansen procedures unrestricted VAR model.
The Johansen procedures inclusion of lagged values of variables, such as income, in its model have little economic plausibility, particularly when using annual data.
Preliminary tests using the Johansen procedure suggest more than one long-term relationship between the relevant variables, and it is difficult to interpret which is the more plausible.
Results from the Engle-Granger approach to cointegration testing of the data are provided in Table 3.8. Due to the non-stationarity
Table 3.8 Cointegration Test Results for South African Cigarette Consumption Data, 19701998
Variable Coefficient T-value Probability Value
Equation 1: Aggregate Annual Cigarette Consumption (Millions of Packs of 20) Constant Real aggregate personal disposable income Real retail price of cigarettes Dummy to neutralize data outlier value for 1982 Dummy for increased tobacco regulation after 1994 Adjusted R = 0.986; DW = 1.633
2
Equation 2: Per Capita Annual Cigarette Consumption (Packs per Annum per Person 15 and over) Constant Real per capita personal disposable income Real retail price of cigarettes Dummy to neutralize data outlier value for 1982 Dummy for increased tobacco regulation after 1994 Adjusted R = 0.922; DW = 1.178
2
85
of the data, the t-values are inflated and should be seen as indicative only. ADF tests are applied to the residuals to test for cointegration in Table 3.9, where the null hypothesis of no cointegration is rejected at the 10 percent level for Equation 3.1 (aggregate cigarette consumption), but not rejected for Equation 3.2 (per capita cigarette consumption). Given that the null hypothesis cannot be confirmed for Equation 3.2, the model is dropped and analysis continued with Equation 3.1.
Johansen Procedure
Once the Johansen cointegration test is applied to estimate the long-run demand relationship, the number of cointegrating equations obtained from the test is then input into the Johansen error correction model. As with the second step of the EngleTable 3.9 Result of ADF Tests for Cointegration on South African Annual Cigarette Consumption Models
Characteristic Trend Intercept Number of lags t-value Critical ADF value at 10% Critical ADF value at 5% Critical ADF value at 1% Equation 1 (Aggregate Consumption) No Yes 2 4.22 4.10 4.50 5.31 Equation 2 (Per Capita Consumption) No Yes 2 3.37 4.10 4.50 5.31
86
Granger method, this provides the parameters of the short-run demand relationships between the variables concerned.
Using instrumental variable techniques to deal with simultaneous determination (see the Apply Instrumental Variable Techniques section of the Specify the Demand Function chapter). Using diagnostic tests (see the Administer Specification and Diagnostic Tests section of the Specify the Demand Function chapter).
87
88
The audience interested in the dissemination of the research results consists of several broadly distinct groups. Their requirements for information from the study differ according to their professional, disciplinary, and functional backgroundsin essence, according to their roles in (or in opposition to) tobacco control in that country.
One of the characteristics differentiating these groups along functional lines is their understanding of, and experience in, economic and econometric analysis. When writing the economic studys research report, take into account the widely differing levels of technical detail at which particular professional groups can (and are willing to) absorb from the results of the study.
It will be the task of the policy makers and analysts concerned with tobacco control, acting in consultation with the coordinators of the overall economic study, to use mass media to communicate the implications of the research results to the wider public.
Economists
This group includes: the overall leader of the economic analysis study, and colleagues working on other economic aspects of tobacco control (e.g., taxation, smuggling, employment, and equity issues)
associates in other organizations, both domestically and overseas, invited to assist with and/or review the analysis (for example, members of the global tobacco control network providing technical support to tobacco control initiatives in their launch phase)
economists in government ministries tasked to examine the research and its implications for economic and other policy (for example, officials of the Budget Office of the Ministry of Finance might be asked to investigate the fiscal implications of the research findings)
89
economists in trade unions organizations, private sector chambers of commerce, industry employer associations, and other bodies representing labor and business, who will have an interest in exploring the wider implications of the research results
Other Professionals
These individuals analyze the implications of tobacco control: health care professionals from the Ministry of Health (tasked with epidemiological monitoring of preventable diseases, preventive program direction, and health systems planning), being permanent civil servants investigating the ways in which economic interventions can support tobacco control initiatives
members of health care NGOs and advocacy groups motivating stronger tobacco control interventions
Policy Makers
These individuals formulate, propagate, debate, and legislate policy measures regarding health and/or economic outcomes: the national and sub-national Ministers of Health, their advisory staffs, and members of health portfolio committees in the national and sub-national legislatures
the national and sub-national Ministers of Finance, their advisory staffs, and members of finance portfolio committees in the national and sub-national legislatures
the national and sub-national Ministers of Labor, their advisory staffs, and members of labor portfolio committees in the national and sub-national legislatures
the national and sub national Ministry of Agriculture or Industry, their advisory staff, and member of agricultural or industry portfolio committees in the national and sub national legislatures, if the county has a major tobacco agricultural or industrial sector
Economists
Most economists have a basic academic knowledge of econometrics. While those conducting the actual research are
90
satisfied with a technical explanation of the demand analysis results, the rest prefer to be presented with a combination of: an overview of the type, frequency, variables, and quality of the data used
a concise discussion of the econometric methodology followed, including a brief review of any specification and/or measurement problems encountered and how these were dealt with a table of the basic regression results for all specifications of the demand model, including comprehensive diagnostic test results
an outline of any further economic research questions raised by the findings of the demand analysis which need to be addressed
Other Professionals
Most professionals of the audience have medical or legal backgrounds. Some have a basic working knowledge of statistics (as opposed to econometrics), but few have more than an undergraduate introduction to economics. Consequently, few professionals are satisfied with a technical explanation of the demand analysis results, and most prefer to receive the following: a brief listing of the type, frequency, variables, and quality of the data used
a concise outline of the econometric methodology followed a table of the regression results and basic diagnostic tests for the demand specification(s) with the most econometrically significant results a table of the estimated elasticities of demand and other basic numerical results, including clear and concise definitions thereof and an appreciation of their degree of precision
a listing of any further economic research questions raised by the findings of the demand analysis which need to be addressed
Policy Makers
Policy makers at all levels are drawn from a wide variety of backgrounds. Except for senior officials from the Ministries of Health and Finance, few have more than an introductory knowledge of statistics, or more than an undergraduate introduction to economics. Policy makers focus much more strongly on the substantive policy implications of the research
91
findings than on the methodology used, as long as they are reasonably assured of the authoritativeness and validity of the research methods. Consequently, policy makers prefer to receive the following information: a brief description of the type of data used, together with an assurance that these data are the best available
a brief description of the econometric methodology followed, together with an assurance that the techniques used are sound and appropriate and that accepted practice is followed in this regard
a table of the estimated elasticities of demand and other basic numerical results, including clear and concise definitions thereof and an appreciation of their degree of precision
a listing of any further economic research questions raised by the findings of the demand analysis which need to be addressed
92
an unfairly high tax burden on the poor increased smuggling of cigarettes into the
country The demand analysis alone cannot produce answers to these issues; instead, input is necessary from the employment, taxation, and equity consideration analyses. Hence it is crucial to coherently present the results of all economic analyses to analysts and policy-makers.
concise discussion of the econometric methodology followed concise discussion of any specification and/or measurement problems encountered and how these were dealt with a table of regression results for all specifications of the demand model, including comprehensive diagnostic test results
a concise discussion of the policy implications of the research results, together with clear flagging of issues to be referred to or cross-referenced with the other economic analyses
93
an outline of any further economic research questions raised by the findings of the demand analysis which need to be addressed
This report forms the basis of the demand analysis component of the Technical Appendix to the external report on the overall economic study.
a description of the broad methodological approach applied to each component a brief description of the nature of the data used for the empirical analyses the values and precision of estimates of key demand factors, and a brief description of the policy implications of these findings a listing of any further economic research questions raised by the findings
2. A concise overview of the research, its results, and policy implications, including:
description of qualitative or quantitative methodology used qualitative and/or quantitative results of analysis policy implications of research results from the viewpoint of the study as a whole
3.
a comprehensive listing of the type, frequency, variables, sources, and quality of the data used concise discussion of the econometric methodology followed
94
concise discussion of any specification and/or measurement problems encountered and how these were dealt with
a table of regression results for all specifications of the demand model, including comprehensive diagnostic test results
Graph the consumption data together with the most significant of the independent variables. This demonstrates the potential impact of particular policy instruments more clearly and with greater impact than words.
Provide an idea of the accuracy of the results (in terms of bias, efficiency, and consistency) and, where necessary, include a sensitivity analysis showing the possible impact of variation in results within the relevant confidence intervals. Spell out in simple terms the implications of estimates being biased versus not being efficient, and so on.
If possible, include a quantitative sensitivity analysis of policy options, explained carefully and simply in text and summarized in table form.
95
X. Additional References
Abedian, I., R. van der Merwe, N. Wilkins, and P. Jha (eds.) (1998), The Economics of Tobacco Control: Towards an Optimal Policy Mix, Cape Town: AFReC, University of Cape Town. Alchin, T.M. (1994), A note on tobacco product prices in the Australian CPI, Applied Economics, Letters, 2:4737. Annett, N. (1996a), An Econometric Estimation of Actual and Potential Government Revenue from Cigarette Taxation in South Africa: 19701995, Project Update No. 2, The Economics of Tobacco Control Project, School of Economics, Cape Town: University of Cape Town. Annett, N. (1996b), An Econometric Analysis of the Effect of Advertising on Cigarette Consumption in South Africa: 19701995, Project Update No. 3, The Economics of Tobacco Control Project, School of Economics, Cape Town: University of Cape Town. Bardsley, P. and N. Olekalns (1999), Cigarette and tobacco consumption: have anti-smoking policies made a difference?, Economic Record, 75(230):22540. Barnett, P.G., T.E. Keeler and T-W. Hu (1995), Oligopoly structure and the incidence of cigarette excise taxes, Journal of Public Economics, 57:45770. Becker, G.S. and K.M. Murphy (1988), A theory of rational addiction, Journal of Political Economy, 96(4):675700. Becker, G.S., M. Grossman and K.M. Murphy (1994), An empirical analysis of cigarette addiction, American Economic Review, 84(3):396418. Bickel, W.K. and G.J. Madden (1999), The behavioural economics of smoking, in Chaloupka, Grossman, Bickel and Saffer (eds.). Bierens, H.J. (1994), Topics in Advanced Econometrics: Estimating, Testing and Specification of Cross-Section and Time Series Models, Cambridge University Press. Bishop, J.A. and J.H. Yoo (1985), Health scare, excise taxes and advertising ban in the cigarette demand and supply, Southern Economic Journal, 40211. Blake Brown, A. (1995), Cigarette taxes and smoking restrictions: impacts and policy implications, American Journal of Agricultural Economics, 77:94651. Chaloupka, F.J. (1991), Rational addictive behaviour and cigarette smoking, Journal of Political Economy, 99(4):72242. Chaloupka, F.J. (1992), Clean indoor air laws, addiction and cigarette smoking, Applied Economics, 24:193205. Chaloupka, F.J. (1999), Macro-social Influences: the Effects of Prices and Tobacco Control Policies on the Demand for Tobacco Products, mimeo. Chaloupka, F.J. and H. Saffer (1992), Clean indoor air laws and the demand for cigarettes, Contemporary Policy Issues, 10:7283.
96
Chaloupka, F.J. and H. Wechsler (1997), Price, tobacco control policies and smoking among young adults, Journal of Health Economics, 16:35973. Chaloupka, F.J. and K.E. Warner (1999a), The Economics of Smoking, Working Paper 7047, National Bureau of Economic Research, Cambridge, MA: NBER. Chaloupka, F.J. and K.E. Warner (1999b), Chapter ?: The economics of smoking, in Jha and Chaloupka (eds.) (2000). Chaloupka, F.J. and M. Grossman (1996), Price, Tobacco Control Policies and Youth Smoking, Working Paper 5740, National Bureau of Economic Research, Cambridge, MA: NBER. Chaloupka, F.J. and R.L. Pacula (1998a), Limiting Youth Access to Tobacco: the Early Impact of the Synar Amendment on Youth Smoking, paper prepared for the 3rd Biennial Pacific Rim Allied Economic Organisations Conference, Bangkok, 14 January. Chaloupka, F.J. and R.L. Pacula (1998b), An Examination of Gender and Race Differences in Youth Smoking Responsiveness to Price and Tobacco Control Policies, Working Paper 6541, National Bureau of Economic Research, Cambridge, MA: NBER. Chaloupka, F.J., J.A. Tauras and M. Grossman (1997), Public policy and youth smokeless tobacco use, Southern Economic Journal, 64(2):50316. Chaloupka, F.J., J.A. Tauras and M. Grossman (1999), Chapter 5: The economics of addiction, in Jha and Chaloupka (eds.) (2000). Chaloupka, F.J., M. Grossman, W.K. Bickel and H. Saffer (1999a), Introduction, in The Economic Analysis of Substance Use and Abuse, Chicago: UC Press. Chaloupka, F.J., P. Jha and R. Peck (1998), A framework for cost-benefit analysis, in Abedian et al. (eds.) (1998). Chaloupka, F.J., R.L. Pacula, M.C. Farrelly, L.D. Johnston, P.M. OMalley and J.W. Bray (1999), Do Higher Cigarette Prices Encourage Youth to Use Marijuana?, Working Paper 6939, National Bureau of Economic Research, Cambridge, MA: NBER. Chaloupka, F.J., T. Hu, K.E. Warner, R. van der Merwe and A. Yurekli (1999b), Chapter 10: The taxation of tobacco products, in Jha and Chaloupka (eds.) (2000). Chapman, S. and J. Richardson (1990), Tobacco excise and declining tobacco consumption: the case of Papua New Guinea, American Journal of Public Health, 80(5):53740. Charemza, W. (1997), New Directions in Econometric Practice: General to Specific Modelling, Cointegration and Vector Autoregression, Edward Elgar. Cragg, J. G. (1971), Some statistical models for limited dependent variables with application to the demand for durable goods, Econometrica, 39(5). Davidson, R. and J.G. MacKinnon (1993), Estimation and Inference in Econometrics, Oxford: OUP. Dougherty, C. (1992), Introduction to Econometrics, Oxford: OUP. Duffy, M. (1995), Advertising in demand systems for alcoholic drinks and tobacco: a comparative study, Journal of Policy Modeling, 17(6):55777. Economics of Tobacco Control in South Africa Project (ETCSA) (1998), The Economics of Tobacco Control in South Africa, report submitted to the International Tobacco Initiative, ETCSA, School of Economics, University of Cape Town. Farrelly, M.C. and J.W. Bray (1998), Response to increases in cigarette prices by race/ethnicity, income, and age groups - United States, 19761993, Morbidity and Mortality Weekly, 47(29):6059.
97
Fujii, E.T. (1980), The demand for cigarettes: further empirical evidence and its implications for public policy, Applied Economics, 12:47989. Greene, W.H. (1991), Econometric Analysis, Prentice-Hall. Grossman, M. and F.J. Chaloupka (1997), Cigarette taxes: the straw to break the camels back, Public Health Reports, 112:291297. Grossman, M., F.J. Chaloupka and R. Anderson (1998), A survey of economic models of addictive behaviour, Journal of Drug Issues, 28(3):63143. Gruber, J. (2000), Youth Smoking in the U.S.: Prices and Policies, Working Paper 7506, National Bureau of Economic Research, Cambridge, MA: NBER. Gujarati, D.N. (1988), Basic Econometrics, Second Edition, McGraw-Hill. Harris, J.E. (1994), A Working Model for Predicting the Consumption and Revenue Impacts of Large Increases in the U.S. Federal Cigarette Excise Tax, mimeo, Cambridge, MA: MIT. Harris, R.I.D. (1995), Using Cointegration Analysis in Econometric Modelling, London: PrenticeHall. Harvey, A. (1990), The Econometric Analysis of Time Series, Second Edition, Philip Allan. Hendry, D.F. (1995), Dynamic Econometrics, Oxford University Press. Hsieh, C., T-W. Hu and C-F.F. Lin (1999), The demand for cigarettes in Taiwan: domestic versus imported cigarettes, Contemporary Economic Policy, 17(2):22334. Hu, T-W. and Y. Tsai (2000), Cigarette Consumption in Rural China: Survey Results from Three Provinces, American Journal of Public Health, 90 (11) 178587. Hu, T-W., H. Sung and T.E. Keeler (1995), The state antismoking campaign and the industry response: the effects of advertising on cigarette consumption in California, American Economic Review, Papers and Proceedings, 85(2):8590. Jha, P. and F.J. Chaloupka (eds.) (2000), Tobacco Control in Developing Countries, published by Oxford University Press on behalf of The World Health Organisation and The World Bank, New York: OUP. World Bank (1999), Curbing the Epidemic: Governments and the Economics of Tobacco Control, Washington, DC: The World Bank. Jha, P., F. Peccaud and S. Nguyen (1999), Chapter 19: Strategic priorities for governments and international agencies in tobacco control, in Jha and Chaloupka (eds.) (2000). Jha, P., P. Musgrove, F.J. Chaloupka and A. Yurekli (1999), Chapter 7: The economic rationale for intervening in the tobacco market, in Jha and Chaloupka (eds.) (2000). Jha, P., T.E. Novotny and R. Feachem (1998), The role of governments in global tobacco control, in Abedian et al. (eds.) (1998). Jimenez, S. and J.M. Labeaga (1994), Is it possible to reduce tobacco consumption via alcohol taxation?, Econometrics and Health Economics, 3:23141. Kao, K. and V.J. Tremblay (1988), Cigarette health scare, excise taxes, and advertising ban: comment, Southern Economic Journal, 7706. Keeler, T., Hu T., P. Barnett and W. Manning (1993), Taxation, regulation and addiction: a demand function for cigarettes based on time-series evidence, Journal of Health Economics, 12:118. Kenkel, D. and L. Chen (1999), Chapter 8: Consumer information and tobacco use, in Jha and Chaloupka (eds.) (2000).
98
Laugesen, M., and C. Meads (1991), Tobacco advertising restrictions, price, income and tobacco consumption in OECD countries, 19601986, British Journal of Addiction, 86(10):134354. Lewit, E.M. and D. Coate (1982), The potential for using excise taxes to reduce smoking, Journal of Health Economics, 1:12145. Lewit, E.M., D. Coate and M. Grossman (1981), The effects of government regulation on teenage smoking, Journal of Law and Economics, 24:54569. Maddala, G.S. (1988), Introduction to Econometrics, New York: Macmillan. Mao, Z., C. Hsieh, T. Hu and J. Jiang (1999), The Demand for Cigarettes in China, mimeo, Institute of Economics, Taipei: Academica Sinica. Mitchell Hoyt, G. and F.J. Chaloupka (1994), Effect of survey conditions on self-reported substance use, Contemporary Economic Policy, 12:10921. Muscatelli, V.A. and S. Hurn (1995), Econometric modelling using cointegrated time series, in Oxley, G., D.A.R. George, C.J. Roberts and S. Sayer (eds), Surveys in Econometrics, Oxford: Blackwell. Novotny, T.E., J.C. Cohen, A. Yurekli, D. Sweanor and J. de Beyer (1999), Chapter 12: Smoking cessation and nicotine replacement therapy, in Jha and Chaloupka (eds.) (2000). Ohsfeldt, R.L., R.G. Boyle and E.I. Capilouto (1999), Tobacco taxes, smoking restrictions, and tobacco use, in Chaloupka, Grossman, Bickel and Saffer (eds.), The Economic Analysis of Substance Use and Abuse, Chicago: UC Press. Saffer, H. (1999), Chapter 9: Tobacco advertising and promotion, in Jha and Chaloupka (eds.) (2000). Saffer, H. and F.J. Chaloupka (1999), Tobacco Advertising: Economic Theory and International Evidence, Working Paper 6958, National Bureau of Economic Research, Cambridge, MA: NBER. Schmalensee, R.L. (1972), On the Economics of Advertising, Amsterdam: North-Holland. Seldon, B.J. and R. Boyd (1991), The stability of cigarette demand, Applied Economics, 23:31926. Stewart, M.J. (1992), Tobacco consumption and advertising restrictions: a critique of Laugesen and Meads (1991), International Journal of Advertising, 11:97118. Stewart, M.J. (1993), The effect on tobacco consumption of advertising bans in OECD countries, International Journal of Advertising, 12:15580. Sung, H., T. Hu and T.E. Keeler (1994), Cigarette taxation and demand: an empirical model, Contemporary Economic Policy, 12:91100. Tauras, J.A. and F.J. Chaloupka (1999a), Price, Clean Indoor Air Laws, and Cigarette Smoking: Evidence from Longitudinal Data for Young Adults, Working Paper 6937, Cambridge, MA: NBER. Tauras, J.A. and F.J. Chaloupka (1999b), Determinants of Smoking Cessation: an Analysis of Young Adult Men and Women, Working Paper 7262, National Bureau of Economic Research, Cambridge, MA: NBER. Townsend, J. (1987), Cigarette tax, economic welfare and social class patterns of smoking, Applied Economics, 19:35565. Townsend, J. (1996), Price and consumption of tobacco, British Medical Bulletin, 52(1):132 42. Townsend, J. (1998), The role of taxation policy in tobacco control, in Abedian et al. (eds.).
99
Townsend, J., P. Roderick and J. Cooper (1994), Cigarette smoking by socio-economic group, sex, and age: effects of price, income and health publicity, British Medical Journal, 309:9237. Tremblay, C.H. and V.J. Tremblay (1995), The impact of cigarette advertising on consumer surplus, profit, and social welfare, Contemporary Economic Policy, 13:11324. Van Walbeek (2000), Impact of the Recent Tobacco Excise Tax Increases on Future Government Revenue Potential, mimeo, The Economics of Tobacco Control in South Africa (Phase II), Applied Fiscal Research Centre, Cape Town: University of Cape Town. Van Walbeek, C.P. (1996), Excise taxes on tobacco: how much scope does the government have?, South African Journal of Economics, 64(1):2042. Warner, K.E. (1977), The effects of the anti-smoking campaign on cigarette consumption, American Journal of Public Health, 67(7):64550. Warner, K.E. (1998), Chapter 6: The economics of health and tobacco: an overview, in Abedian, et al. (eds.). Warner, K.E., F.J. Chaloupka, P.J. Cook, W.G. Manning, J.P. Newhouse, T.E. Novotny, T.C. Schelling and J. Townsend (1995), Criteria for determining an optimal cigarette tax: the economists perspective, Tobacco Control, 4:3806. Wasserman, J., W.E. Manning, J.P. Newhouse and J.D. Winkler (1991), The effects of excise taxes and regulation on cigarette smoking, Journal of Health Economics, 1:4364. Willig, R.D. (1976), Consumers surplus without apology, American Economic Review, 66(4):58997. Woollery, T., S. Asma and D. Sharp (1999), Chapter 11: Clean indoor laws and youth access restrictions, in Jha and Chaloupka (eds.) (2000). Xu, Hu and Keeler (2000), Optimal Cigarette Taxation: Theory and Estimation, mimeo.
100